diff --git a/.gitignore b/.gitignore
index 64c415c..af1ac9e 100644
--- a/.gitignore
+++ b/.gitignore
@@ -215,4 +215,6 @@ __marimo__/
 # Streamlit
 .streamlit/secrets.toml
 
+.kilo/
+logs/
 progress.json
diff --git a/core/objective.py b/core/objective.py
index cf15d9e..5851514 100644
--- a/core/objective.py
+++ b/core/objective.py
@@ -1,3 +1,6 @@
+from dataclasses import dataclass
+from typing import Optional
+
 import random
 
 from scipy.ndimage import map_coordinates
@@ -9,50 +12,58 @@ from core.cylinder import generate_cylinder_n_torch, generate_cylinder_o_torch,
 from core.intersection import center_line_intersections_torch
 from core.scoring import cl_score_torch, cl_score_torch_xfr
 
-# Global variables (used in objective_function)
-image1_array = None      # cortical_nii.gz
-image2_array = None      # binarynii.gz
-image2_shape = None
-image3_array = None      # roi2.nii.gz
-diameter = None
-length = None
-spacing = [0.5, 0.5, 0.5]
-device = None        
-grid = None 
-USE_TIP_PENALTY = None
 
-def set_global_context(
-    cortical,
-    spine,
-    shape,
-    spacing_,
-    device_,
-    grid_,
-    use_tip_penalty=False  # 新增
-):
-    global cortical_tensor, spine_tensor, image2_shape, spacing, device, grid, USE_TIP_PENALTY
+# =====================================================================
+# OptimizationContext: the single explicit container for all shared state
+# needed to evaluate a candidate cylinder. New code should build one of
+# these and bind it to an objective via make_objective_function[_xfr];
+# no module globals are read during optimization.
+# =====================================================================
+
+@dataclass
+class OptimizationContext:
+    """Shared state for evaluating candidate cylinder placements.
+
+    Required:
+        cortical_tensor: cortical bone mask (uint8, 0/1)
+        spine_tensor:    bone (binary) mask (uint8, 0/1)
+        image2_shape:    (Z, Y, X) volume shape
+        spacing:         voxel spacing, [sx, sy, sz] (mm)
+        device:          torch device for tensor ops
+
+    Optional:
+        grid:            precomputed coordinate grid (z_t, y_t, x_t)
+        use_tip_penalty: add the tip-cylinder penalty to the loss
+        The *_array / spine_roi_tensor / diameter / length fields are kept
+        for compatibility and debugging; the loss itself does not use them.
+    """
+    cortical_tensor: torch.Tensor
+    spine_tensor: torch.Tensor
+    image2_shape: tuple
+    spacing: list
+    device: torch.device
+    grid: Optional[tuple] = None
+    use_tip_penalty: bool = False
+    spine_roi_tensor: Optional[torch.Tensor] = None
+    image1_array: Optional[np.ndarray] = None
+    image2_array: Optional[np.ndarray] = None
+    image3_array: Optional[np.ndarray] = None
+    diameter: Optional[float] = None
+    length: Optional[float] = None
 
-    cortical_tensor = cortical
-    spine_tensor = spine
-    image2_shape = shape
-    spacing = spacing_
-    device = device_
-    grid = grid_
-    USE_TIP_PENALTY = use_tip_penalty
 
 def cylinder_circle_line_intersection_loss_deductions_torch(
+    ctx: OptimizationContext,
     diameter: float,
     length: float,
     params: list[float],
-    image_shape: tuple[int, int, int],
-    cortical_tensor: torch.Tensor,
-    spine_tensor: torch.Tensor,
-    spacing: list[float],
-    device: torch.device
 ) -> float:
     """
-    Computes the loss for a given set of cylinder params in PyTorch, 
+    Computes the loss for a given set of cylinder params in PyTorch,
     returning a Python float for PSO consumption.
+
+    All shared state (tensors, shape, spacing, device, grid, tip penalty)
+    is taken from `ctx`; no module globals are involved.
     """
     position_z, position_y, position_x, azimuth, altitude = params
 
@@ -64,10 +75,10 @@ def cylinder_circle_line_intersection_loss_deductions_torch(
         position_x,
         float(azimuth),
         float(altitude),
-        image_shape,
-        spacing,
-        device,
-        grid
+        ctx.image2_shape,
+        ctx.spacing,
+        ctx.device,
+        ctx.grid
     )
 
     cyl_opp = generate_cylinder_o_torch(
@@ -78,13 +89,12 @@ def cylinder_circle_line_intersection_loss_deductions_torch(
         position_x,
         float(azimuth),
         float(altitude),
-        image_shape,
-        spacing,
-        device,
-        grid
+        ctx.image2_shape,
+        ctx.spacing,
+        ctx.device,
+        ctx.grid
     )
 
-
     # We call the center_line_intersections in Torch mode
     intersections, _ = center_line_intersections_torch(
         position_z,
@@ -93,89 +103,200 @@ def cylinder_circle_line_intersection_loss_deductions_torch(
         azimuth,
         altitude,
         length,
-        spine_tensor,
-        spacing,
-        device
+        ctx.spine_tensor,
+        ctx.spacing,
+        ctx.device
     )
 
     cyl_tip = None
-    if USE_TIP_PENALTY:
+    if ctx.use_tip_penalty:
         cyl_tip = generate_cylinder_tip_torch(
             diameter, length,
             position_z, position_y, position_x,
             float(azimuth), float(altitude),
-            image_shape, spacing, device, grid
+            ctx.image2_shape, ctx.spacing, ctx.device, ctx.grid
         )
 
     # loss_value = cl_score_torch(
     loss_value = cl_score_torch_xfr(
-        cortical_tensor, spine_tensor, 
+        ctx.cortical_tensor, ctx.spine_tensor,
         cyl_fwd, cyl_opp, intersections,
         cylinder_tip_torch=cyl_tip
     )
 
     return loss_value
 
-def objective_function_xfr(params: list[float], y_indices) -> float:
+
+# =====================================================================
+# Context-bound objective builders (preferred API)
+# =====================================================================
+
+def _evaluate(params: list[float], ctx: OptimizationContext) -> float:
     """
-    Wrapper for the PSO objective function, calling our Torch-based loss function.
-    Now params includes diameter and length at the end.
-    params = [position_z, position_y, position_x, azimuth, altitude, diameter_raw, length_raw]
-    """
-
-    # position_params = params[:5]  # [z, y, x, azimuth, altitude]
-    # diameter_raw = params[5]
-    # length_raw = params[6]
-
-    z, x, azimuth, altitude, diameter_raw, length_raw = params
-    y = y_indices[round(z), round(x)] #+ random.uniform(-0.5, 0.5)
-    
-    # coords = np.array([[z], [x]])
-    # result = map_coordinates(y_indices, coords, order=1)
-    # y= result[0]
-
-    position_params = [z, y, x, azimuth, altitude]
-
-    # 將連續值轉換為離散值
-    # diameter_discrete, length_discrete = snap_to_discrete_values(diameter_raw, length_raw)
-    diameter_discrete, length_discrete = snap_to_discrete_values_xfr(diameter_raw, length_raw)
-
-    diameter_loss = .9*diameter_discrete + .1*diameter_raw
-    length_loss   = .9*  length_discrete + .1*  length_raw
-    
-    loss = cylinder_circle_line_intersection_loss_deductions_torch(
-        diameter_loss,
-        length_loss,
-        position_params,
-        image2_shape,
-        cortical_tensor,
-        spine_tensor,
-        spacing,
-        device
-    )
-    return loss
-
-def objective_function(params: list[float]) -> float:
-    """
-    Wrapper for the PSO objective function, calling our Torch-based loss function.
-    Now params includes diameter and length at the end.
-    params = [position_z, position_y, position_x, azimuth, altitude, diameter_raw, length_raw]
+    Core objective: params = [z, y, x, azimuth, altitude, diameter_raw, length_raw]
     """
     position_params = params[:5]  # [z, y, x, azimuth, altitude]
     diameter_raw = params[5]
     length_raw = params[6]
-    
+
     # 將連續值轉換為離散值
     diameter_discrete, length_discrete = snap_to_discrete_values(diameter_raw, length_raw)
-    
-    loss = cylinder_circle_line_intersection_loss_deductions_torch(
+
+    return cylinder_circle_line_intersection_loss_deductions_torch(
+        ctx,
         diameter_discrete,
         length_discrete,
-        position_params,
-        image2_shape,
-        cortical_tensor,
-        spine_tensor,
-        spacing,
-        device
+        position_params
     )
-    return loss
\ No newline at end of file
+
+
+def _evaluate_xfr(params: list[float], ctx: OptimizationContext, y_indices) -> float:
+    """
+    Core xfr objective: params = [z, x, azimuth, altitude, diameter_raw, length_raw];
+    y is derived from the per-column `y_indices` surface.
+    """
+    z, x, azimuth, altitude, diameter_raw, length_raw = params
+    y = y_indices[round(z), round(x)] #+ random.uniform(-0.5, 0.5)
+
+    position_params = [z, y, x, azimuth, altitude]
+
+    # 將連續值轉換為離散值
+    diameter_discrete, length_discrete = snap_to_discrete_values_xfr(diameter_raw, length_raw)
+
+    diameter_loss = .9*diameter_discrete + .1*diameter_raw
+    length_loss   = .9*  length_discrete + .1*  length_raw
+
+    return cylinder_circle_line_intersection_loss_deductions_torch(
+        ctx,
+        diameter_loss,
+        length_loss,
+        position_params
+    )
+
+
+def make_objective_function(ctx: OptimizationContext):
+    """
+    Return an objective function bound to `ctx`.
+    Pass the returned callable directly to pso / differential_evolution / minimize.
+    """
+    def objective(params: list[float]) -> float:
+        return _evaluate(params, ctx)
+    return objective
+
+
+def make_objective_function_xfr(ctx: OptimizationContext, y_indices):
+    """
+    Return an xfr objective function bound to `ctx` and the `y_indices` surface.
+    Pass the returned callable directly to pso / differential_evolution / minimize.
+    """
+    def objective(params: list[float]) -> float:
+        return _evaluate_xfr(params, ctx, y_indices)
+    return objective
+
+
+# =====================================================================
+# Legacy API (backward compatible)
+#
+# The old flow mutated module attributes on this file from the optimizers
+# ("跨檔案注入變數"), which was fragile: any forgotten attribute showed up
+# deep inside an optimizer callback as a NoneType/NameError. It is kept so
+# existing callers (set_global_context + objective_function) keep working,
+# but new code should use OptimizationContext + the make_* factories.
+# =====================================================================
+
+# Module-level state for the legacy path only.
+cortical_tensor = None
+spine_tensor = None
+image1_array = None      # cortical_nii.gz
+image2_array = None      # binarynii.gz
+image2_shape = None
+image3_array = None      # roi2.nii.gz
+diameter = None
+length = None
+spacing = [0.5, 0.5, 0.5]
+device = None
+grid = None
+USE_TIP_PENALTY = None
+
+_current_context: Optional[OptimizationContext] = None
+
+
+def set_global_context(
+    cortical,
+    spine,
+    shape,
+    spacing_,
+    device_,
+    grid_,
+    use_tip_penalty=False
+):
+    """
+    Legacy: set the process-wide context used by the legacy
+    objective_function / objective_function_xfr wrappers.
+
+    Returns the built OptimizationContext for convenience.
+    """
+    global _current_context
+    global cortical_tensor, spine_tensor, image2_shape, spacing, device, grid, USE_TIP_PENALTY
+
+    _current_context = OptimizationContext(
+        cortical_tensor=cortical,
+        spine_tensor=spine,
+        image2_shape=shape,
+        spacing=spacing_,
+        device=device_,
+        grid=grid_,
+        use_tip_penalty=use_tip_penalty,
+    )
+
+    # mirror into legacy module attributes for any code that reads them
+    cortical_tensor = cortical
+    spine_tensor = spine
+    image2_shape = shape
+    spacing = spacing_
+    device = device_
+    grid = grid_
+    USE_TIP_PENALTY = use_tip_penalty
+
+    return _current_context
+
+
+def _active_context() -> OptimizationContext:
+    """Resolve the context for the legacy wrappers, with a clear error."""
+    if _current_context is not None:
+        return _current_context
+
+    # Fallback: a caller (old-style optimizers) may have set the legacy
+    # module attributes directly — rebuild a context from them.
+    if cortical_tensor is not None and spine_tensor is not None:
+        return OptimizationContext(
+            cortical_tensor=cortical_tensor,
+            spine_tensor=spine_tensor,
+            image2_shape=image2_shape,
+            spacing=spacing,
+            device=device,
+            grid=grid,
+            use_tip_penalty=bool(USE_TIP_PENALTY),
+        )
+
+    raise RuntimeError(
+        "No optimization context available. Either call set_global_context(...) "
+        "or (preferred) build an OptimizationContext and use "
+        "make_objective_function / make_objective_function_xfr."
+    )
+
+
+def objective_function(params: list[float]) -> float:
+    """
+    Legacy wrapper: evaluates against the context set by set_global_context.
+    params = [z, y, x, azimuth, altitude, diameter_raw, length_raw]
+    """
+    return _evaluate(params, _active_context())
+
+
+def objective_function_xfr(params: list[float], y_indices) -> float:
+    """
+    Legacy wrapper: evaluates against the context set by set_global_context.
+    params = [z, x, azimuth, altitude, diameter_raw, length_raw]
+    """
+    return _evaluate_xfr(params, _active_context(), y_indices)
\ No newline at end of file
diff --git a/core/optimizer.py b/core/optimizer.py
index eea5f7b..1d08cd1 100644
--- a/core/optimizer.py
+++ b/core/optimizer.py
@@ -2,16 +2,89 @@ import time
 from datetime import datetime
 import SimpleITK as sitk
 import torch
-from imaging.orientation import azimuth_rotation, analyze_vertebral_tilt_contour
+from imaging.orientation import (azimuth_rotation, analyze_vertebral_tilt_contour,
+                                  best_symmetry_plane, best_upper_endplate_plane,
+                                  segment_spinous_process)
 from config.constant import ALLOWED_DIAMETERS, ALLOWED_LENGTHS
-from core.objective import objective_function, objective_function_xfr
-from pyswarm import pso 
-import core.objective  # <--- 加入這行，讓我們可以直接操作 objective 模組
-from core.cylinder import generate_cylinder_n_torch, snap_to_discrete_values, create_coordinate_grid, snap_to_discrete_values_xfr
-from core.scoring import compute_overlap_ratio_from_cylinder_mask, is_solution_ok
+from core.objective import OptimizationContext, make_objective_function, make_objective_function_xfr
+from pyswarm import pso
+from core.cylinder import generate_cylinder_n_torch, generate_cylinder_o_torch, snap_to_discrete_values, create_coordinate_grid, snap_to_discrete_values_xfr
+from core.intersection import center_line_intersections_torch
+from core.scoring import compute_overlap_ratio_from_cylinder_mask, is_solution_ok, cl_score_torch_xfr
 from config.constant import OVERLAP_THRESH
 from visualization.res_plot_3d import res_plt_2_torch
 
+LATERAL_REFINE_MIN_IN_BONE = 0.97
+
+
+def refine_lateral_longer(
+    z, x, azimuth, altitude, diameter_raw, length_raw,
+    side, az_bounds, x_bounds, y_indices,
+    image_shape, spacing, device, grid,
+    cortical_tensor, spine_tensor,
+):
+    """
+    Deterministic local refinement after PSO: try aiming more laterally and
+    using a longer screw. Only accepts a candidate if it stays
+    >= LATERAL_REFINE_MIN_IN_BONE inside bone AND improves the score.
+
+    left  (x-lower half): more lateral = larger azimuth, entry shifted toward -x
+    right (x-upper half): more lateral = smaller azimuth, entry shifted toward +x
+    """
+    def _score_candidate(cand):
+        z_c, x_c, az_c, alt_c, d_c, L_c = cand
+        y_c = y_indices[round(z_c), round(x_c)]
+        if y_c < 0:
+            return None
+        cyl = generate_cylinder_n_torch(d_c, L_c, z_c, y_c, x_c, az_c, alt_c,
+                                        image_shape, spacing, device, grid)
+        cyl_o = generate_cylinder_o_torch(d_c, L_c, z_c, y_c, x_c, az_c, alt_c,
+                                          image_shape, spacing, device, grid)
+        if cyl.sum().item() == 0:
+            return None
+        inter, _ = center_line_intersections_torch(z_c, y_c, x_c, az_c, alt_c,
+                                                   L_c, spine_tensor, spacing, device)
+        loss = cl_score_torch_xfr(cortical_tensor, spine_tensor, cyl, cyl_o, inter)
+        in_bone = ((spine_tensor == 1) & (cyl == 1)).sum().item() / cyl.sum().item()
+        return {'pos': cand, 'loss': loss, 'in_bone': in_bone}
+
+    d_snap, L_snap = snap_to_discrete_values_xfr(diameter_raw, length_raw)
+    az_lo, az_hi = az_bounds
+    x_lo, x_hi = x_bounds
+
+    az_steps = [0.0, 5.0, 10.0, 15.0] if side == "L" else [0.0, -5.0, -10.0, -15.0]
+    x_shifts = [0.0, -4.0, -8.0] if side == "L" else [0.0, 4.0, 8.0]
+
+    best = _score_candidate((z, x, azimuth, altitude, d_snap, L_snap))
+    if best is None:
+        return z, x, azimuth, altitude, diameter_raw, length_raw, False, None
+
+    for ds in az_steps:
+        az_c = min(az_hi - 0.01, max(az_lo + 0.01, azimuth + ds))
+        if abs(az_c - azimuth) < 0.5 and ds != 0:
+            continue
+        for dx in x_shifts:
+            x_c = min(x_hi, max(x_lo, x + dx))
+            if abs(x_c - x) < 0.5 and dx != 0:
+                continue
+            for L_c in sorted({L_snap} | {l for l in ALLOWED_LENGTHS if l > L_snap}):
+                cand = _score_candidate((z, x_c, az_c, altitude, d_snap, L_c))
+                if cand is None:
+                    continue
+                if cand['in_bone'] >= LATERAL_REFINE_MIN_IN_BONE and cand['loss'] < best['loss']:
+                    best = cand
+
+    z_r, x_r, az_r, alt_r, d_r, L_r = best['pos']
+    adopted = not (abs(z_r - z) < 1e-6 and abs(x_r - x) < 1e-6
+                   and abs(az_r - azimuth) < 1e-6 and abs(L_r - L_snap) < 1e-6)
+    return z_r, x_r, az_r, alt_r, d_r, L_r, adopted, best
+
+
+def _validate_bounds(lb, ub, name):
+    for i, (lo, hi) in enumerate(zip(lb, ub)):
+        if lo >= hi:
+            raise ValueError(f'PSO bounds invalid for {name}: dim {i} lower {lo} >= upper {hi}')
+
 
 def get_first_nonzero_y(arr):
     OUTSIDE_VALUE = -100
@@ -80,59 +153,100 @@ def run_pso_torch_xfr(
     spine_tensor = torch.from_numpy(image2_array).to(device=device, dtype=torch.uint8)
     spine_roi_tensor = torch.from_numpy(image3_array).to(device=device, dtype=torch.uint8)
 
-    # ================= [跨檔案注入變數：終極防呆版] =================
-    import core.objective
-
-    # 1. 注入 Tensors
-    core.objective.cortical_tensor = cortical_tensor
-    core.objective.spine_tensor = spine_tensor
-    core.objective.spine_roi_tensor = spine_roi_tensor
-
-    # 2. 注入 Arrays (以防 objective 裡面偷偷用到 Numpy 陣列)
-    core.objective.image1_array = image1_array
-    core.objective.image2_array = image2_array
-    core.objective.image3_array = image3_array
-
-    # 3. 注入 Shapes (這就是導致這次 NoneType 報錯的真兇！)
-    core.objective.image2_shape = image2_shape  # <--- 解除警報的最關鍵一行
-    core.objective.image_shape = image_shape
-    core.objective.shape = image_shape
-
-    # 4. 注入環境變數
-    core.objective.spacing = spacing
-    core.objective.device = device
-    core.objective.grid = grid
-
-    # 5. 注入尺寸參數 (兼容固定尺寸模式)
-    if not optimize_size:
-        core.objective.diameter = diameter
-        core.objective.length = length
-    # ==============================================================
+    # 建立明確的優化上下文（取代舊的跨檔案 global 注入；狀態全部顯式傳入目標函數）
+    ctx = OptimizationContext(
+        cortical_tensor=cortical_tensor,
+        spine_tensor=spine_tensor,
+        spine_roi_tensor=spine_roi_tensor,
+        image1_array=image1_array,
+        image2_array=image2_array,
+        image3_array=image3_array,
+        image2_shape=image2_shape,
+        spacing=spacing,
+        device=device,
+        grid=grid,
+        diameter=diameter if not optimize_size else None,
+        length=length if not optimize_size else None,
+    )
 
     azi = azimuth_rotation(image2_path)
     res = analyze_vertebral_tilt_contour(image2_path, edge_type='superior', show_plot=False, debug=False)
     alt = res['superior']['tilt_angle_deg']
 
+    # ===== az/alt 搜尋範圍錨點：以椎體自身座標系取代 2D 近似 =====
+    # 鏡稱對稱面（最佳 mirror plane）：法線 (s_nx, s_ny, s_nz)，s_nx>0 固定符號
+    #   theta_v = atan2(s_ny, s_nx) 是椎體真實左右軸相對 +x 的旋轉角，
+    #   取代舊 2D 輪廓前點角 azi（舊式 az 中心 90-azi == 新式 90+theta_v）。
+    # 上終板面（normal 朝上, e_nz>0）：
+    #   tau_y = atan2(e_ny, e_nz)：終板 AP 面傾斜；負 = 面向 +y（終板側）升高，
+    #       軌跡需爬得更陡 → altitude 中心減小（舊式 65-alt 的 3D 版）。
+    #   tau_x = atan2(e_nx, e_nz)：終板 LR 面傾斜；正 = 面向小 x（L 側）升高，
+    #       L 側需更陡（-tau_x）、R 側較平（+tau_x）。
+    sym_plane = best_symmetry_plane(image2_array)
+
+    # 入口面 y_indices 取「每個 (z,x) 柱第一個 bone voxel（最後側）」，
+    # 中線柱會落在棘突上（入口太靠內後）。先把棘突（鏡稱面中線後側，
+    # 見 segment_spinous_process）從 image2_array 移除再取 surface，
+    # 讓 y_indices 永不落在棘突上（loss 用 spine_tensor，不受影響）。
+    sp_mask, sp_th, sp_info = segment_spinous_process(image2_array, sym_plane)
+    if sp_mask is not None and sp_mask.any():
+        n_sp = int(sp_mask.sum())
+        image2_array[sp_mask] = 0
+        print(f"[Y-INDEX] spinous process removed from entry surface: {n_sp} vox "
+              f"(band=+/-{sp_info['band_w']:.1f} voxel, AP>={sp_info['ap_thresh']:.1f}, "
+              f"mode={sp_info['mode']})")
+
     y_indices = get_first_nonzero_y(image2_array)
+    objective_fn = make_objective_function_xfr(ctx, y_indices)
+
+    s_nx, s_ny, s_nz = sym_plane['normal']
+    theta_v = float(np.degrees(np.arctan2(s_ny, s_nx)))
+    endplate_plane = best_upper_endplate_plane(image2_array)
+    if endplate_plane is not None:
+        e_nx, e_ny, e_nz = endplate_plane['normal']
+        tau_y = float(np.degrees(np.arctan2(e_ny, e_nz)))
+        tau_x = float(np.degrees(np.arctan2(e_nx, e_nz)))
+    else:
+        tau_y, tau_x = 0.0, 0.0
+    print(f"[PLANE] mirror  : {sym_plane['plane'][0]:+.3f}x {sym_plane['plane'][1]:+.3f}y "
+          f"{sym_plane['plane'][2]:+.3f}z = {sym_plane['plane'][3]:.1f}"
+          f"  (theta_v={theta_v:+.2f} deg, mirror ratio={sym_plane['ratio']:.3f})")
+    if endplate_plane is not None:
+        p = endplate_plane['plane']
+        print(f"[PLANE] endplate: {p[0]:+.3f}x {p[1]:+.3f}y {p[2]:+.3f}z = {p[3]:.1f}"
+              f"  (tau_y={tau_y:+.2f} deg, tau_x={tau_x:+.2f} deg, "
+              f"inlier={endplate_plane['inlier_ratio']:.2f})")
+    else:
+        print("[PLANE] endplate: 資料不足，altitude 錨點退回 tau_y=tau_x=0")
 
     # flat_min_index = np.argmin(y_indices)
     # z_border, x_border = np.unravel_index(flat_min_index, y_indices.shape)
 
-    x_with_nonzero = np.where(np.any(image2_array[:,image_shape[1]//10,:] != 0, axis=0))[0]
+    # 脊椎中線：整段 (全體積) 骨頭 x 範圍的中點。
+    # 不取單一行的原因 (0005 L5)：椎體軸狀面旋轉時單行只罩到單側骨塊
+    # (x 1..76 / W=224 → x_mid≈0.17W)，L/R 兩個 band 被壓到同一側。
+    # 不用鏡稱對稱軸的原因 (0001 L4)：逐切面對稱軸會被肋、後側要素
+    # 左右不對稱與椎體傾斜牽引 (69.0 vs 範圍中點 74.5)，把 R band 內緣
+    # (x_mid+0.1W) 拉進中線棘突/椎板區，R 側入口落在棘突上 (太靠內後)。
+    # Laminectomy 只移除中線後側要素，左右極端 x 位置不變，
+    # 所以範圍中點同樣不受其影響，作為 L/R band 分割線比對稱軸穩定。
+    x_with_nonzero = np.where(np.any(image2_array != 0, axis=(0, 1)))[0]
     x1 = x_with_nonzero[0]
     x2 = x_with_nonzero[-1]
+    x_width = x2 - x1
 
     # print(x1,x2)
     # exit()
 
-    x_mid = (x1+x2)/2
-    x1 = x_mid-image_shape[2]*.1
-    x2 = x_mid+image_shape[2]*.1
+    # x_mid = (x1 + x2) / 2
+    # x1 = x_mid-image_shape[2]*.1
+    # x2 = x_mid+image_shape[2]*.1
 
     z_sum = np.sum(image2_array, axis=(1, 2))
     z_with_nonzero = np.where(z_sum > 0)[0]
     z1 = z_with_nonzero[0]
     z2 = z_with_nonzero[-1]   
+    z_height = z2-z1
 
     # print(x1,x2)
     # exit()
@@ -149,18 +263,37 @@ def run_pso_torch_xfr(
     if CBT == True:    
         # z_bounds = (0, image_shape[0]-1)
         # z_bounds = (z1, (z1+z2)/2)
-        z_bounds = (.1*image_shape[0], .8*image_shape[0])
+        # z_bounds = (.1*image_shape[0], .8*image_shape[0])
+        z_bounds = (z1+z_height*.1, z1+z_height*.9)
+
         # x_bounds_right = (image_shape[2]/2 + image_shape[2]/10, image_shape[2] - 1)
         # x_bounds_left  = (0, image_shape[2]/2 - image_shape[2]/10 - 1)
-        x_bounds_right = (x2, image_shape[2]*.9)
-        x_bounds_left  = (image_shape[2]*.1, x1)
+        # x_bounds_right = (x2, image_shape[2]*.9)
+        # x_bounds_left  = (image_shape[2]*.1, x1)
+        x_bounds_right = (x1+x_width*.6, +x_width*.9)
+        x_bounds_left  = (x1+x_width*.1, +x_width*.4)
 
-        # azimuth_bounds_l = ((95-azi), (145-azi))
-        # azimuth_bounds_r = ((50-azi), (85-azi))
-        # altitude_bounds = ((60-alt), (75-alt))
-        azimuth_bounds_l = ((98-azi), (120-azi))
-        azimuth_bounds_r = ((60-azi), (82-azi))
-        altitude_bounds = ((60-alt), (70-alt))
+        # 脊椎若被體積邊界切到（真正偏心、骨頭貼著左/右邊緣），
+        # 對應那側的 x band 下限會 >= 上限，PSO 會丟 "upper-bound must be greater"。
+        # 出錯時 clamp 成同側最小寬度（5% 寬度）的合法 band。
+        min_band = .05 * image_shape[2]
+        if x_bounds_left[1] <= x_bounds_left[0]:
+            x_bounds_left = (x_bounds_left[0], x_bounds_left[0] + min_band)
+        if x_bounds_right[0] >= x_bounds_right[1]:
+            x_bounds_right = (x_bounds_right[1] - min_band, x_bounds_right[1])
+
+        # 舊 2D 版本（以輪廓角 azi / 矢狀面傾斜 alt 平移固定範圍），保留供對照：
+        # azimuth_bounds_l = ((98-azi), (120-azi))
+        # azimuth_bounds_r = ((60-azi), (82-azi))
+        # altitude_bounds = ((60-alt), (70-alt))
+        # 新：範圍以「椎體自身座標系」為中心 —— 椎體系中 az=90° 是 AP 直向、
+        # L 帶 = AP 後退 8~30°（偏 -x）、R 帶 = AP 前進 8~30°（偏 +x）、
+        # altitude 60~70。再換算回影像系：az 整體加 theta_v（鏡稱面），
+        # altitude 加 tau_y（終板 AP 傾斜）並逐側加 -/+tau_x（終板 LR 傾斜）。
+        azimuth_bounds_l = ((98+theta_v), (120+theta_v))
+        azimuth_bounds_r = ((60+theta_v), (82+theta_v))
+        altitude_bounds_l = ((60+tau_y-tau_x), (70+tau_y-tau_x))
+        altitude_bounds_r = ((60+tau_y+tau_x), (70+tau_y+tau_x))
 
     else:      
         z_bounds = (0, image_shape[0] - 1)   
@@ -170,6 +303,9 @@ def run_pso_torch_xfr(
         azimuth_bounds_l = (60-azi, 90-azi) 
         azimuth_bounds_r = (90-azi, 120-azi) 
         altitude_bounds = (65-alt, 80-alt)
+        # 非 CBT 分支維持單一式，兩側同用
+        altitude_bounds_l = altitude_bounds
+        altitude_bounds_r = altitude_bounds
 
     def eval_overlap_from_position(pos, side: str, optimize_size: bool,
                                spine_tensor: torch.Tensor,
@@ -204,16 +340,16 @@ def run_pso_torch_xfr(
         diameter_bounds = (min(ALLOWED_DIAMETERS), max(ALLOWED_DIAMETERS)*1.01)
         length_bounds = (min(ALLOWED_LENGTHS), max(ALLOWED_LENGTHS)*1.01)
         
-        # bounds 現在有 7 個參數
+        # bounds 現在有 7 個參數（altitude 分 L/R 兩側，由終板面 tau_x 決定）
         lb_l = [z_bounds[0], x_bounds_left[0], azimuth_bounds_l[0], 
-                altitude_bounds[0], diameter_bounds[0], length_bounds[0]]
+                altitude_bounds_l[0], diameter_bounds[0], length_bounds[0]]
         ub_l = [z_bounds[1], x_bounds_left[1], azimuth_bounds_l[1], 
-                altitude_bounds[1], diameter_bounds[1], length_bounds[1]]
+                altitude_bounds_l[1], diameter_bounds[1], length_bounds[1]]
 
         lb_r = [z_bounds[0], x_bounds_right[0], azimuth_bounds_r[0], 
-                altitude_bounds[0], diameter_bounds[0], length_bounds[0]]
+                altitude_bounds_r[0], diameter_bounds[0], length_bounds[0]]
         ub_r = [z_bounds[1], x_bounds_right[1], azimuth_bounds_r[1], 
-                altitude_bounds[1], diameter_bounds[1], length_bounds[1]]
+                altitude_bounds_r[1], diameter_bounds[1], length_bounds[1]]
         
     else:
         # 模式 2：固定 diameter 和 length（向後兼容）
@@ -222,11 +358,11 @@ def run_pso_torch_xfr(
         diameter = 4.5  # 或從參數傳入
         length = 45     # 或從參數傳入
         
-        lb_l = [z_bounds[0], y_bounds[0], x_bounds_left[0], azimuth_bounds_l[0], altitude_bounds[0]]
-        ub_l = [z_bounds[1], y_bounds[1], x_bounds_left[1], azimuth_bounds_l[1], altitude_bounds[1]]
+        lb_l = [z_bounds[0], y_bounds[0], x_bounds_left[0], azimuth_bounds_l[0], altitude_bounds_l[0]]
+        ub_l = [z_bounds[1], y_bounds[1], x_bounds_left[1], azimuth_bounds_l[1], altitude_bounds_l[1]]
 
-        lb_r = [z_bounds[0], y_bounds[0], x_bounds_right[0], azimuth_bounds_r[0], altitude_bounds[0]]
-        ub_r = [z_bounds[1], y_bounds[1], x_bounds_right[1], azimuth_bounds_r[1], altitude_bounds[1]]
+        lb_r = [z_bounds[0], y_bounds[0], x_bounds_right[0], azimuth_bounds_r[0], altitude_bounds_r[0]]
+        ub_r = [z_bounds[1], y_bounds[1], x_bounds_right[1], azimuth_bounds_r[1], altitude_bounds_r[1]]
 
     if True or debug:
         print(lb_l)
@@ -241,17 +377,31 @@ def run_pso_torch_xfr(
 
     # Left side optimization
     print(f"\n=== {label_str} 左側 ===")
-    kwargs = {'y_indices': y_indices}
-    position_l, loss_l = pso(objective_function_xfr, lb_l, ub_l,
-    
+    _validate_bounds(lb_l, ub_l, f'{label_str} L')
+    position_l, loss_l = pso(objective_fn, lb_l, ub_l,
         # ieqcons=[constraint_y],
-        kwargs=kwargs,    
-        swarmsize=swarm_size, 
+        swarmsize=swarm_size,
         omega = omega,
         maxiter=max_iter, debug=debug)
 
     z, x, azimuth, altitude, diameter, length = position_l
+    az_pso, x_pso, L_pso = azimuth, x, length
     y = y_indices[round(z), round(x)]
+
+    z, x, azimuth, altitude, diameter, length, adopted_l, ref_l = refine_lateral_longer(
+        z, x, azimuth, altitude, diameter, length,
+        "L", azimuth_bounds_l, x_bounds_left, y_indices,
+        image_shape, spacing, device, grid,
+        cortical_tensor, spine_tensor,
+    )
+    y = y_indices[round(z), round(x)]
+    if adopted_l:
+        loss_l = ref_l['loss']
+        print(f"[LEFT] lateral-refine: az {az_pso:.2f} -> {azimuth:.2f}, x {x_pso:.2f} -> {x:.2f}, "
+              f"L {L_pso:.2f} -> {length:.2f}, in-bone {ref_l['in_bone']*100:.1f}%")
+    else:
+        print("[LEFT] lateral-refine: no improvement")
+
     position_l = z, y, x, azimuth, altitude, diameter, length
 
     overlap_l, diameter_l, length_l = eval_overlap_from_position(
@@ -277,7 +427,7 @@ def run_pso_torch_xfr(
 
     # 左側 retry：loss 要 <=0 且 overlap >= 0.5 才算過關
     # while (best_loss_l > 0 or best_overlap_l < OVERLAP_THRESH) and retries < max_retries:
-        # position_l, loss_l = pso(objective_function, lb_l, ub_l, swarmsize=swarm_size, maxiter=max_iter)
+        # position_l, loss_l = pso(objective_fn, lb_l, ub_l, swarmsize=swarm_size, maxiter=max_iter)
         # overlap_l, diameter_l, length_l = eval_overlap_from_position(
             # position_l, "L", optimize_size, spine_tensor, image_shape, spacing
         # )
@@ -307,15 +457,31 @@ def run_pso_torch_xfr(
     
     # Right side optimization
     print(f"\n=== {label_str} 右側 ===")
-    position_r, loss_r = pso(objective_function_xfr, lb_r, ub_r,
+    _validate_bounds(lb_r, ub_r, f'{label_str} R')
+    position_r, loss_r = pso(objective_fn, lb_r, ub_r,
         # ieqcons=[constraint_y],
-        kwargs=kwargs,    
         swarmsize=swarm_size,
         omega = omega,
         maxiter=max_iter, debug=debug)
 
     z, x, azimuth, altitude, diameter, length = position_r
+    az_pso, x_pso, L_pso = azimuth, x, length
     y = y_indices[round(z), round(x)]
+
+    z, x, azimuth, altitude, diameter, length, adopted_r, ref_r = refine_lateral_longer(
+        z, x, azimuth, altitude, diameter, length,
+        "R", azimuth_bounds_r, x_bounds_right, y_indices,
+        image_shape, spacing, device, grid,
+        cortical_tensor, spine_tensor,
+    )
+    y = y_indices[round(z), round(x)]
+    if adopted_r:
+        loss_r = ref_r['loss']
+        print(f"[RIGHT] lateral-refine: az {az_pso:.2f} -> {azimuth:.2f}, x {x_pso:.2f} -> {x:.2f}, "
+              f"L {L_pso:.2f} -> {length:.2f}, in-bone {ref_r['in_bone']*100:.1f}%")
+    else:
+        print("[RIGHT] lateral-refine: no improvement")
+
     position_r = z, y, x, azimuth, altitude, diameter, length
 
     overlap_r, diameter_r, length_r = eval_overlap_from_position(
@@ -345,7 +511,7 @@ def run_pso_torch_xfr(
     # retries = 0
 
     # while (best_loss_r > 0 or best_overlap_r < OVERLAP_THRESH) and retries < max_retries:
-        # position_r, loss_r = pso(objective_function, lb_r, ub_r, swarmsize=swarm_size, maxiter=max_iter)
+        # position_r, loss_r = pso(objective_fn, lb_r, ub_r, swarmsize=swarm_size, maxiter=max_iter)
         # overlap_r, diameter_r, length_r = eval_overlap_from_position(
             # position_r, "R", optimize_size, spine_tensor, image_shape, spacing
         # )
@@ -458,34 +624,22 @@ def run_pso_torch(
     spine_tensor = torch.from_numpy(image2_array).to(device=device, dtype=torch.uint8)
     spine_roi_tensor = torch.from_numpy(image3_array).to(device=device, dtype=torch.uint8)
 
-    # ================= [跨檔案注入變數：終極防呆版] =================
-    import core.objective
-
-    # 1. 注入 Tensors
-    core.objective.cortical_tensor = cortical_tensor
-    core.objective.spine_tensor = spine_tensor
-    core.objective.spine_roi_tensor = spine_roi_tensor
-
-    # 2. 注入 Arrays (以防 objective 裡面偷偷用到 Numpy 陣列)
-    core.objective.image1_array = image1_array
-    core.objective.image2_array = image2_array
-    core.objective.image3_array = image3_array
-
-    # 3. 注入 Shapes (這就是導致這次 NoneType 報錯的真兇！)
-    core.objective.image2_shape = image2_shape  # <--- 解除警報的最關鍵一行
-    core.objective.image_shape = image_shape
-    core.objective.shape = image_shape
-
-    # 4. 注入環境變數
-    core.objective.spacing = spacing
-    core.objective.device = device
-    core.objective.grid = grid
-
-    # 5. 注入尺寸參數 (兼容固定尺寸模式)
-    if not optimize_size:
-        core.objective.diameter = diameter
-        core.objective.length = length
-    # ==============================================================
+    # 建立明確的優化上下文（取代舊的跨檔案 global 注入；狀態全部顯式傳入目標函數）
+    ctx = OptimizationContext(
+        cortical_tensor=cortical_tensor,
+        spine_tensor=spine_tensor,
+        spine_roi_tensor=spine_roi_tensor,
+        image1_array=image1_array,
+        image2_array=image2_array,
+        image3_array=image3_array,
+        image2_shape=image2_shape,
+        spacing=spacing,
+        device=device,
+        grid=grid,
+        diameter=diameter if not optimize_size else None,
+        length=length if not optimize_size else None,
+    )
+    objective_fn = make_objective_function(ctx)
 
     azi = azimuth_rotation(image2_path)
     res = analyze_vertebral_tilt_contour(image2_path, edge_type='superior', show_plot=False, debug=False)
@@ -497,9 +651,11 @@ def run_pso_torch(
         y_bounds = (image_shape[1]/5, image_shape[1]/2 - 1)
         x_bounds_right = (image_shape[2]/2 + image_shape[2]/10, image_shape[2] - 1)
         x_bounds_left  = (0, image_shape[2]/2 - image_shape[2]/10 - 1)
-        azimuth_bounds_l = ((95-azi), (145-azi))
-        azimuth_bounds_r = ((50-azi), (85-azi))
-        altitude_bounds = ((60-alt), (75-alt))
+        # CBT 參數依據 references/：Santoni 2009 (Spine J 9:366) 冠狀面 25-30° cranial (caudo-cephalad)、
+        # 軸狀面自正中線向外 (medial→lateral) ≤30°；Delgado-Fernandez 2017 (ASJ 11:817)、Kim 2022 (SSRR 6:1)
+        azimuth_bounds_l = ((98-azi), (120-azi))
+        azimuth_bounds_r = ((60-azi), (82-azi))
+        altitude_bounds = ((60-alt), (70-alt))
 
         # xfr
         # z_bounds = (0, image_shape[0] - 1)
@@ -588,7 +744,7 @@ def run_pso_torch(
 
     # Left side optimization
     print("\n=== 左側 ===")
-    position_l, loss_l = pso(objective_function, lb_l, ub_l, swarmsize=swarm_size, maxiter=max_iter, debug=debug)
+    position_l, loss_l = pso(objective_fn, lb_l, ub_l, swarmsize=swarm_size, maxiter=max_iter, debug=debug)
 
     overlap_l, diameter_l, length_l = eval_overlap_from_position(
         position_l, "L", optimize_size, spine_tensor, image_shape, spacing
@@ -612,7 +768,7 @@ def run_pso_torch(
 
     # 左側 retry：loss 要 <=0 且 overlap >= 0.5 才算過關
     # while (best_loss_l > 0 or best_overlap_l < OVERLAP_THRESH) and retries < max_retries:
-        # position_l, loss_l = pso(objective_function, lb_l, ub_l, swarmsize=swarm_size, maxiter=max_iter)
+        # position_l, loss_l = pso(objective_fn, lb_l, ub_l, swarmsize=swarm_size, maxiter=max_iter)
         # overlap_l, diameter_l, length_l = eval_overlap_from_position(
             # position_l, "L", optimize_size, spine_tensor, image_shape, spacing
         # )
@@ -642,7 +798,7 @@ def run_pso_torch(
     
     # Right side optimization
     print("\n=== 右側 ===")
-    position_r, loss_r = pso(objective_function, lb_r, ub_r, swarmsize=swarm_size, maxiter=max_iter, debug=debug)
+    position_r, loss_r = pso(objective_fn, lb_r, ub_r, swarmsize=swarm_size, maxiter=max_iter, debug=debug)
     overlap_r, diameter_r, length_r = eval_overlap_from_position(
         position_r, "R", optimize_size, spine_tensor, image_shape, spacing
     )
@@ -670,7 +826,7 @@ def run_pso_torch(
     # retries = 0
 
     # while (best_loss_r > 0 or best_overlap_r < OVERLAP_THRESH) and retries < max_retries:
-        # position_r, loss_r = pso(objective_function, lb_r, ub_r, swarmsize=swarm_size, maxiter=max_iter)
+        # position_r, loss_r = pso(objective_fn, lb_r, ub_r, swarmsize=swarm_size, maxiter=max_iter)
         # overlap_r, diameter_r, length_r = eval_overlap_from_position(
             # position_r, "R", optimize_size, spine_tensor, image_shape, spacing
         # )
@@ -748,7 +904,6 @@ from scipy.optimize import differential_evolution
 from scipy.optimize import minimize
 from imaging.orientation import azimuth_rotation, analyze_vertebral_tilt_contour
 from config.constant import ALLOWED_DIAMETERS, ALLOWED_LENGTHS
-from core.objective import objective_function
 from core.cylinder import generate_cylinder_n_torch, snap_to_discrete_values, create_coordinate_grid
 from core.scoring import compute_overlap_ratio_from_cylinder_mask, is_solution_ok
 from config.constant import OVERLAP_THRESH
@@ -790,34 +945,22 @@ def run_de_torch(
     spine_tensor = torch.from_numpy(image2_array).to(device=device, dtype=torch.uint8)
     spine_roi_tensor = torch.from_numpy(image3_array).to(device=device, dtype=torch.uint8)
 
-    # ================= [跨檔案注入變數：終極防呆版] =================
-    import core.objective
-
-    # 1. 注入 Tensors
-    core.objective.cortical_tensor = cortical_tensor
-    core.objective.spine_tensor = spine_tensor
-    core.objective.spine_roi_tensor = spine_roi_tensor
-
-    # 2. 注入 Arrays (以防 objective 裡面偷偷用到 Numpy 陣列)
-    core.objective.image1_array = image1_array
-    core.objective.image2_array = image2_array
-    core.objective.image3_array = image3_array
-
-    # 3. 注入 Shapes (這就是導致這次 NoneType 報錯的真兇！)
-    core.objective.image2_shape = image2_shape  # <--- 解除警報的最關鍵一行
-    core.objective.image_shape = image_shape
-    core.objective.shape = image_shape
-
-    # 4. 注入環境變數
-    core.objective.spacing = spacing
-    core.objective.device = device
-    core.objective.grid = grid
-
-    # 5. 注入尺寸參數 (兼容固定尺寸模式)
-    if not optimize_size:
-        core.objective.diameter = diameter
-        core.objective.length = length
-    # ==============================================================
+    # 建立明確的優化上下文（取代舊的跨檔案 global 注入；狀態全部顯式傳入目標函數）
+    ctx = OptimizationContext(
+        cortical_tensor=cortical_tensor,
+        spine_tensor=spine_tensor,
+        spine_roi_tensor=spine_roi_tensor,
+        image1_array=image1_array,
+        image2_array=image2_array,
+        image3_array=image3_array,
+        image2_shape=image2_shape,
+        spacing=spacing,
+        device=device,
+        grid=grid,
+        diameter=diameter if not optimize_size else None,
+        length=length if not optimize_size else None,
+    )
+    objective_fn = make_objective_function(ctx)
 
     azi = azimuth_rotation(image2_path)
     res = analyze_vertebral_tilt_contour(image2_path, edge_type='superior', show_plot=False, debug=False)
@@ -828,9 +971,11 @@ def run_de_torch(
         y_bounds = (image_shape[1]/5, image_shape[1]/2 - 1)
         x_bounds_right = (image_shape[2]/2 + image_shape[2]/10, image_shape[2] - 1)
         x_bounds_left  = (0, image_shape[2]/2 - image_shape[2]/10 - 1)
-        azimuth_bounds_l = ((95-azi), (145-azi))
-        azimuth_bounds_r = ((50-azi), (85-azi))
-        altitude_bounds = ((60-alt), (75-alt))
+        # CBT 參數依據 references/：Santoni 2009 (Spine J 9:366) 冠狀面 25-30° cranial (caudo-cephalad)、
+        # 軸狀面自正中線向外 (medial→lateral) ≤30°；Delgado-Fernandez 2017 (ASJ 11:817)、Kim 2022 (SSRR 6:1)
+        azimuth_bounds_l = ((98-azi), (120-azi))
+        azimuth_bounds_r = ((60-azi), (82-azi))
+        altitude_bounds = ((60-alt), (70-alt))
     else:      
         z_bounds = (0, image_shape[0] - 1)   
         y_bounds = (image_shape[1]/5, image_shape[1]/2 - 1)   
@@ -877,7 +1022,7 @@ def run_de_torch(
 
     # --- 左側最佳化 ---
     print("\n=== 左側 (DE) ===")
-    res_l = differential_evolution(objective_function, bounds_l, popsize=de_popsize, maxiter=max_iter)
+    res_l = differential_evolution(objective_fn, bounds_l, popsize=de_popsize, maxiter=max_iter)
     position_l, loss_l = res_l.x, res_l.fun
 
     overlap_l, diameter_l, length_l = eval_overlap_from_position(position_l, "L", optimize_size, spine_tensor, image_shape, spacing)
@@ -886,7 +1031,7 @@ def run_de_torch(
     """
     retries = 0
     while (best_loss_l > 0 or best_overlap_l < OVERLAP_THRESH) and retries < 10:
-        res_l = differential_evolution(objective_function, bounds_l, popsize=de_popsize, maxiter=max_iter)
+        res_l = differential_evolution(objective_fn, bounds_l, popsize=de_popsize, maxiter=max_iter)
         position_l, loss_l = res_l.x, res_l.fun
         overlap_l, diameter_l, length_l = eval_overlap_from_position(position_l, "L", optimize_size, spine_tensor, image_shape, spacing)
         
@@ -899,7 +1044,7 @@ def run_de_torch(
     """
     # --- 右側最佳化 ---
     print("\n=== 右側 (DE) ===")
-    res_r = differential_evolution(objective_function, bounds_r, popsize=de_popsize, maxiter=max_iter)
+    res_r = differential_evolution(objective_fn, bounds_r, popsize=de_popsize, maxiter=max_iter)
     position_r, loss_r = res_r.x, res_r.fun
 
     overlap_r, diameter_r, length_r = eval_overlap_from_position(position_r, "R", optimize_size, spine_tensor, image_shape, spacing)
@@ -908,7 +1053,7 @@ def run_de_torch(
     """
     retries = 0
     while (best_loss_r > 0 or best_overlap_r < OVERLAP_THRESH) and retries < 10:
-        res_r = differential_evolution(objective_function, bounds_r, popsize=de_popsize, maxiter=max_iter)
+        res_r = differential_evolution(objective_fn, bounds_r, popsize=de_popsize, maxiter=max_iter)
         position_r, loss_r = res_r.x, res_r.fun
         overlap_r, diameter_r, length_r = eval_overlap_from_position(position_r, "R", optimize_size, spine_tensor, image_shape, spacing)
         
@@ -969,34 +1114,22 @@ def run_nm_torch(
     cortical_tensor = torch.from_numpy(image1_array).to(device=device, dtype=torch.uint8)
     spine_tensor = torch.from_numpy(image2_array).to(device=device, dtype=torch.uint8)
     spine_roi_tensor = torch.from_numpy(image3_array).to(device=device, dtype=torch.uint8)
-    # ================= [跨檔案注入變數：終極防呆版] =================
-    import core.objective
-
-    # 1. 注入 Tensors
-    core.objective.cortical_tensor = cortical_tensor
-    core.objective.spine_tensor = spine_tensor
-    core.objective.spine_roi_tensor = spine_roi_tensor
-
-    # 2. 注入 Arrays (以防 objective 裡面偷偷用到 Numpy 陣列)
-    core.objective.image1_array = image1_array
-    core.objective.image2_array = image2_array
-    core.objective.image3_array = image3_array
-
-    # 3. 注入 Shapes (這就是導致這次 NoneType 報錯的真兇！)
-    core.objective.image2_shape = image2_shape  # <--- 解除警報的最關鍵一行
-    core.objective.image_shape = image_shape
-    core.objective.shape = image_shape
-
-    # 4. 注入環境變數
-    core.objective.spacing = spacing
-    core.objective.device = device
-    core.objective.grid = grid
-
-    # 5. 注入尺寸參數 (兼容固定尺寸模式)
-    if not optimize_size:
-        core.objective.diameter = diameter
-        core.objective.length = length
-    # ==============================================================
+    # 建立明確的優化上下文（取代舊的跨檔案 global 注入；狀態全部顯式傳入目標函數）
+    ctx = OptimizationContext(
+        cortical_tensor=cortical_tensor,
+        spine_tensor=spine_tensor,
+        spine_roi_tensor=spine_roi_tensor,
+        image1_array=image1_array,
+        image2_array=image2_array,
+        image3_array=image3_array,
+        image2_shape=image2_shape,
+        spacing=spacing,
+        device=device,
+        grid=grid,
+        diameter=diameter if not optimize_size else None,
+        length=length if not optimize_size else None,
+    )
+    objective_fn = make_objective_function(ctx)
 
     azi = azimuth_rotation(image2_path)
     res = analyze_vertebral_tilt_contour(image2_path, edge_type='superior', show_plot=False, debug=False)
@@ -1007,9 +1140,11 @@ def run_nm_torch(
         y_bounds = (image_shape[1]/5, image_shape[1]/2 - 1)
         x_bounds_right = (image_shape[2]/2 + image_shape[2]/10, image_shape[2] - 1)
         x_bounds_left  = (0, image_shape[2]/2 - image_shape[2]/10 - 1)
-        azimuth_bounds_l = ((95-azi), (145-azi))
-        azimuth_bounds_r = ((50-azi), (85-azi))
-        altitude_bounds = ((60-alt), (75-alt))
+        # CBT 參數依據 references/：Santoni 2009 (Spine J 9:366) 冠狀面 25-30° cranial (caudo-cephalad)、
+        # 軸狀面自正中線向外 (medial→lateral) ≤30°；Delgado-Fernandez 2017 (ASJ 11:817)、Kim 2022 (SSRR 6:1)
+        azimuth_bounds_l = ((98-azi), (120-azi))
+        azimuth_bounds_r = ((60-azi), (82-azi))
+        altitude_bounds = ((60-alt), (70-alt))
     else:      
         z_bounds = (0, image_shape[0] - 1)   
         y_bounds = (image_shape[1]/5, image_shape[1]/2 - 1)   
@@ -1054,7 +1189,7 @@ def run_nm_torch(
     # --- 左側最佳化 ---
     print("\n=== 左側 (Nelder-Mead) ===")
     x0_l = get_random_x0(bounds_l)
-    res_l = minimize(objective_function, x0_l, method='Nelder-Mead', bounds=bounds_l, options={'maxiter': max_iter})
+    res_l = minimize(objective_fn, x0_l, method='Nelder-Mead', bounds=bounds_l, options={'maxiter': max_iter})
     position_l, loss_l = res_l.x, res_l.fun
 
     overlap_l, diameter_l, length_l = eval_overlap_from_position(position_l, "L", optimize_size, spine_tensor, image_shape, spacing)
@@ -1064,7 +1199,7 @@ def run_nm_torch(
     retries = 0
     while (best_loss_l > 0 or best_overlap_l < OVERLAP_THRESH) and retries < 10:
         x0_l = get_random_x0(bounds_l) # 每次 retry 都換一個隨機起始點
-        res_l = minimize(objective_function, x0_l, method='Nelder-Mead', bounds=bounds_l, options={'maxiter': max_iter})
+        res_l = minimize(objective_fn, x0_l, method='Nelder-Mead', bounds=bounds_l, options={'maxiter': max_iter})
         position_l, loss_l = res_l.x, res_l.fun
         overlap_l, diameter_l, length_l = eval_overlap_from_position(position_l, "L", optimize_size, spine_tensor, image_shape, spacing)
         
@@ -1078,7 +1213,7 @@ def run_nm_torch(
     # --- 右側最佳化 ---
     print("\n=== 右側 (Nelder-Mead) ===")
     x0_r = get_random_x0(bounds_r)
-    res_r = minimize(objective_function, x0_r, method='Nelder-Mead', bounds=bounds_r, options={'maxiter': max_iter})
+    res_r = minimize(objective_fn, x0_r, method='Nelder-Mead', bounds=bounds_r, options={'maxiter': max_iter})
     position_r, loss_r = res_r.x, res_r.fun
 
     overlap_r, diameter_r, length_r = eval_overlap_from_position(position_r, "R", optimize_size, spine_tensor, image_shape, spacing)
@@ -1088,7 +1223,7 @@ def run_nm_torch(
     retries = 0
     while (best_loss_r > 0 or best_overlap_r < OVERLAP_THRESH) and retries < 10:
         x0_r = get_random_x0(bounds_r)
-        res_r = minimize(objective_function, x0_r, method='Nelder-Mead', bounds=bounds_r, options={'maxiter': max_iter})
+        res_r = minimize(objective_fn, x0_r, method='Nelder-Mead', bounds=bounds_r, options={'maxiter': max_iter})
         position_r, loss_r = res_r.x, res_r.fun
         overlap_r, diameter_r, length_r = eval_overlap_from_position(position_r, "R", optimize_size, spine_tensor, image_shape, spacing)
         
diff --git a/core/scoring.py b/core/scoring.py
index 14d39c1..a4679fd 100644
--- a/core/scoring.py
+++ b/core/scoring.py
@@ -36,10 +36,12 @@ def cl_score_torch_xfr(
     # if in_bone == 0:
     #     return float(not_in_bone*200)
 
-    score += 10   * in_bone # 10 實在太低
+    score += 20   * in_bone # 10 實在太低
     score += 100  * overlap
-    score -= 2000 * max(0, not_in_bone-10)
-    score -= 1000 * max(0, null_vox2-10)
+    # score -= 2000 * max(0, not_in_bone-10)
+    # score -= 1000 * max(0, null_vox2-10)
+    score -= 1000 * not_in_bone
+    score -= 1000 * null_vox2
 
     return float(-score)
     
diff --git a/imaging/orientation.py b/imaging/orientation.py
index 6f75892..eae41df 100644
--- a/imaging/orientation.py
+++ b/imaging/orientation.py
@@ -1,8 +1,309 @@
 import numpy as np
 import SimpleITK as sitk
-from scipy.ndimage import center_of_mass
+from scipy.ndimage import center_of_mass, rotate
 import matplotlib.pyplot as plt
 
+def _best_vertical_split(mask2d):
+    """
+    binary 2D 陣列（最後一軸 = x）：找讓 mask 與其鏡射重疊最大的垂直線 x = t。
+    c[s] = Σ_u A[u]·A[s-u] 是每列自卷積；批次 FFT 一次算出所有 s。
+    回傳 (score, s_best)；s = 2·t（t 可能為半整數）。
+    """
+    n = mask2d.shape[-1]
+    flat = np.asarray(mask2d).reshape(-1, n)
+    rows = flat[flat.max(axis=1) > 0]
+    if rows.size == 0:
+        return 0.0, 0
+    L = 1 << (2 * n - 1).bit_length()
+    padded = np.zeros((rows.shape[0], L), dtype=np.float32)
+    padded[:, :n] = rows.astype(np.float32)
+    F = np.fft.rfft(padded, axis=1)
+    conv = np.fft.irfft(F * F, axis=1)[:, :2 * n - 1]
+    score = np.clip(conv.sum(axis=0), 0, None)
+    s_best = int(np.argmax(score))
+    return float(score[s_best]), s_best
+
+
+def best_symmetry_axis_angle(proj, coarse_step=10.0, fine_step=1.0):
+    """
+    2D binary 投影（axis0=row y, axis1=col x）的左右鏡稱軸搜尋：
+    將影像旋轉 θ 後，鏡稱軸變成垂直線（col = const），用 _best_vertical_split 評分；
+    粗搜 0..180°（coarse_step）再在 winner 附近細搜（fine_step）。
+    回傳 (theta_deg, score, s_best)。
+    """
+    base = (np.asarray(proj) > 0).astype(np.float32)
+    best_th, best, best_s = 0.0, -1.0, 0
+    for th in np.arange(0.0, 180.0, coarse_step):
+        sc, s = _best_vertical_split(rotate(base, th, reshape=False, order=0) > 0.5)
+        if sc > best:
+            best_th, best, best_s = float(th), sc, s
+    for th in np.arange(best_th - coarse_step, best_th + coarse_step, fine_step):
+        th2 = float(th) % 180.0
+        sc, s = _best_vertical_split(rotate(base, th2, reshape=False, order=0) > 0.5)
+        if sc > best:
+            best_th, best, best_s = th2, sc, s
+    return best_th, best, best_s
+
+
+def _mirror_hit_count(X, Y, Z, n, d, m, sz, sy, sx):
+    """每個骨 voxel 對平面 n·p=d 鏡射後四捨五入到最近 voxel，回傳命中骨 voxel 的數量"""
+    nx, ny, nz = n
+    dist = X * nx + Y * ny + Z * nz - d
+    rx = (X - 2.0 * dist * nx).round().astype(np.int32)
+    ry = (Y - 2.0 * dist * ny).round().astype(np.int32)
+    rz = (Z - 2.0 * dist * nz).round().astype(np.int32)
+    ok = (rx >= 0) & (rx < sx) & (ry >= 0) & (ry < sy) & (rz >= 0) & (rz < sz)
+    if not ok.any():
+        return 0
+    return int(m[rz[ok], ry[ok], rx[ok]].sum())
+
+
+def best_symmetry_plane(mask_zyx, phi_max=45.0, subsample=7,
+                        coarse_step=7.5, fine_step=1.0):
+    """
+    3D bone mask (z, y, x) 的最佳鏡稱面，一般平面方程 a·x + b·y + c·z = d
+    （voxel index 座標；(a,b,c) 為單位法線，方向任意，不限制平行 YZ 面）。
+    以「每個 bone voxel 鏡射後四捨五入到最近 voxel 的命中率」為分數，
+    參數化 n = R_y(phi)·R_z(theta)·(1,0,0)：
+      theta : 法線在 axial (x,y) 平面內的旋轉
+      phi   : 法線出平面的傾斜，限制在 [-phi_max, +phi_max] 以確保仍切分左右
+    搜尋：theta 由 2D axial 投影粗定位，再 3D 三段式（粗→細→微細）
+    （粗/細階段用等距子樣 voxel 加速）。
+    回傳 dict：
+      plane     : (a, b, c, d)  ->  a*x + b*y + c*z = d
+      normal    : (a, b, c)
+      offset    : d
+      theta_deg / phi_deg : 法線參數
+      ratio     : 鏡射命中率（0..1）
+      u, v      : 平面內兩個正交方向（供繪製用）
+    """
+    m = np.asarray(mask_zyx) > 0
+    sz, sy, sx = m.shape
+    zz, yy, xx = np.nonzero(m)
+    if xx.size < 100:
+        n = np.array([1.0, 0.0, 0.0])
+        d = (sx - 1) / 2.0
+        return {'plane': (1.0, 0.0, 0.0, float(d)), 'normal': (1.0, 0.0, 0.0),
+                'offset': float(d), 'theta_deg': 0.0, 'phi_deg': 0.0,
+                'ratio': 1.0, 'u': (0.0, 1.0, 0.0), 'v': (0.0, 0.0, 1.0)}
+    xf = xx.astype(np.float32)
+    yf = yy.astype(np.float32)
+    zf = zz.astype(np.float32)
+    N = xf.size
+    c0 = np.array([(sx - 1) / 2.0, (sy - 1) / 2.0, (sz - 1) / 2.0])
+    Xs, Ys, Zs = xf[::subsample], yf[::subsample], zf[::subsample]
+
+    def n_of(theta_deg, phi_deg):
+        t = np.deg2rad(theta_deg)
+        p = np.deg2rad(phi_deg)
+        return np.array([np.cos(t) * np.cos(p), np.sin(t), -np.cos(t) * np.sin(p)])
+
+    # theta 初值：2D axial 投影搜尋（垂直面情形）
+    proj = m.max(axis=0)
+    theta0 = best_symmetry_axis_angle(proj.astype(np.float32))[0] if proj.sum() >= 10 else 0.0
+
+    best = None
+
+    def consider(theta, phi, d, full=False):
+        nonlocal best
+        n = n_of(theta, phi)
+        X, Y, Z = (xf, yf, zf) if full else (Xs, Ys, Zs)
+        s = _mirror_hit_count(X, Y, Z, n, d, m, sz, sy, sx)
+        if best is None or s > best['score']:
+            best = {'score': s, 'theta': float(theta), 'phi': float(phi), 'd': float(d)}
+
+    # stage 1：粗搜尋（子樣）
+    for th in np.arange(theta0 - coarse_step * 2, theta0 + coarse_step * 2 + 1e-9, coarse_step):
+        for ph in np.arange(-phi_max, phi_max + 1e-9, coarse_step):
+            n = n_of(th, ph)
+            d0 = float(n @ c0)
+            for dd in (-10.0, 0.0, 10.0):
+                consider(th, ph, d0 + dd)
+    # stage 2：細搜尋（子樣）
+    for th in np.arange(best['theta'] - 2 * fine_step * 2, best['theta'] + 2 * fine_step * 2 + 1e-9, fine_step):
+        for ph in np.arange(best['phi'] - 2 * fine_step * 2, best['phi'] + 2 * fine_step * 2 + 1e-9, fine_step):
+            for dd in (-3.0, -1.0, 0.0, 1.0, 3.0):
+                consider(th, ph, best['d'] + dd)
+    # stage 3：微細搜尋（全分辨率）
+    for th in np.arange(best['theta'] - fine_step, best['theta'] + fine_step + 1e-9, fine_step / 2):
+        for ph in np.arange(best['phi'] - fine_step, best['phi'] + fine_step + 1e-9, fine_step / 2):
+            for dd in (-1.0, -0.5, 0.0, 0.5, 1.0):
+                consider(th, ph, best['d'] + dd, full=True)
+    # stage 4：d 微調（全分辨率）
+    b = best
+    for dd in np.arange(-1.0, 1.0 + 1e-9, 0.25):
+        consider(b['theta'], b['phi'], b['d'] + dd, full=True)
+
+    n = n_of(best['theta'], best['phi'])
+    theta = best['theta']
+    d = best['d']
+    if theta > 90.0 or theta < -90.0:     # 正規化到 (-90, 90]，法線翻轉時 d 取反
+        theta -= 180.0
+        n = -n
+        d = -d
+    ratio = _mirror_hit_count(xf, yf, zf, n, d, m, sz, sy, sx) / N
+    u = np.cross(n, [0.0, 0.0, 1.0])
+    if np.linalg.norm(u) < 0.1:
+        u = np.cross(n, [1.0, 0.0, 0.0])
+    u = u / np.linalg.norm(u)
+    v = np.cross(n, u)
+    return {
+        'plane': (float(n[0]), float(n[1]), float(n[2]), float(d)),
+        'normal': (float(n[0]), float(n[1]), float(n[2])),
+        'offset': float(d),
+        'theta_deg': float(theta),
+        'phi_deg': float(best['phi']),
+        'ratio': float(ratio),
+        'u': (float(u[0]), float(u[1]), float(u[2])),
+        'v': (float(v[0]), float(v[1]), float(v[2])),
+    }
+
+def segment_spinous_process(mask_zyx, sym, band_frac=0.06, min_band=6.0,
+                            min_mass_frac=0.05):
+    """
+    以 best_symmetry_plane 的結果 sym 從 3D bone mask (z, y, x) 切出棘突。
+    棘突是中線後側構造，利用鏡稱面 a·x+b·y+c·z=d 定義：
+      1) 中線帶：骨 voxel 到平面的有號距離 |s| <= w，
+         w = max(min_band, band_frac * s 全寬)
+      2) 前後方向：平面內兩軸 (u, v) 中 |y| 分量大者，
+         正規化成 +AP = 後側（本資料系 y 往前遞增，後側 = y 小側）
+      3) 中線帶的 AP 分佈呈兩大叢（椎體在前、椎弓/棘突在後），
+         以兩叢間的 AP 谷底為界，AP >= 谷底 的中線帶 voxel = 棘突（含中線椎弓）；
+         無明顯谷底（如骨橋）fallback 取中線帶後側 15%。
+    回傳 (sp_mask (z,y,x) bool, ap_thresh, info dict)；
+    資料過少時 sp_mask = None（info['mode'] 說明原因）。
+    """
+    m = np.asarray(mask_zyx) > 0
+    zz, yy, xx = np.nonzero(m)
+    info = {'n_bone': int(zz.size), 'n_sp': 0, 'ap_thresh': None,
+            'band_w': None, 'mode': 'empty'}
+    if zz.size < 50:
+        return None, None, info
+    a, b, c, d = sym['plane']
+    X = xx.astype(np.float64)
+    Y = yy.astype(np.float64)
+    Z = zz.astype(np.float64)
+    s = X * a + Y * b + Z * c - d
+    u = np.array(sym['u'])
+    v = np.array(sym['v'])
+    u_ap = u if abs(u[1]) >= abs(v[1]) else v
+    if u_ap[1] > 0:
+        u_ap = -u_ap                                  # +AP = 後側（y 小側）
+    w = max(float(min_band), float(band_frac) * float(s.max() - s.min()))
+    mid = np.abs(s) <= w
+    ap = X * u_ap[0] + Y * u_ap[1] + Z * u_ap[2]
+    aps = ap[mid]
+    info['band_w'] = float(w)
+    if aps.size < 50:
+        info['mode'] = 'too_few_midline'
+        return None, None, info
+    lo = int(np.floor(aps.min()))
+    hi = int(np.ceil(aps.max()))
+    th = None
+    mode = 'fallback'
+    if hi - lo >= 10:
+        hist, edges = np.histogram(aps, bins=range(lo, hi + 1))
+        csum = np.concatenate([[0], np.cumsum(hist)])
+        total = csum[-1]
+        peak = hist.max()
+        best_i, best_score = None, -1.0
+        for i in range(len(hist)):
+            if hist[i] >= 0.05 * peak:
+                continue
+            if csum[i] < min_mass_frac * total or (total - csum[i + 1]) < min_mass_frac * total:
+                continue
+            score = min(csum[i], total - csum[i + 1])
+            if score > best_score:
+                best_score, best_i = score, i
+        if best_i is not None:
+            th = float(0.5 * (edges[best_i] + edges[best_i + 1]))
+            mode = 'gap'
+    if th is None:
+        th = float(np.quantile(aps, 0.85))
+    sp_mask = np.zeros(m.shape, dtype=bool)
+    sel = mid & (ap >= th)
+    sp_mask[zz[sel], yy[sel], xx[sel]] = True
+    info.update(n_sp=int(sel.sum()), ap_thresh=th, mode=mode)
+    return sp_mask, th, info
+
+def best_upper_endplate_plane(mask_zyx, angle_max=45.0, thresh=3.0,
+                              n_iter=500, seed=42):
+    """
+    3D bone mask (z, y, x) 的最佳「上終板」近似平面 a·x + b·y + c·z = d
+    （voxel index 座標；(a,b,c) 為朝上的單位法線）。
+    1) 每個 (y, x) 欄位取最上方 bone voxel 作為頂面點（僅前側半邊，
+       y >= COM_y，避開後方元素，與 2D superior endplate 定義一致）
+    2) RANSAC 三點擬平面：法線限制在與 +z 軸 ≤ angle_max° 內，
+       計數 ±thresh voxel 內的頂面點為 inlier，取 inlier 最多者
+    3) SVD 最小二乘微調
+    回傳 dict（資料不足時回傳 None）：
+      plane : (a, b, c, d)
+      normal / offset
+      tilt_deg     : 法線與 +z 軸的夾角（上終板傾斜）
+      inlier_ratio : 頂面點落在平面 ±thresh 的比例
+      n_points / n_inliers
+      u, v : 平面內正交方向（供繪製用）
+    """
+    m = np.asarray(mask_zyx) > 0
+    nz, ny, nx = m.shape
+    idx = np.where(m, np.arange(nz)[:, None, None], -1)
+    ztop = idx.max(axis=0)                                # (y, x) 每欄最上 z
+    y_split = int(round(center_of_mass(m)[1])) if m.sum() else 0
+    # ztop 是 (y, x)：條件作用在 y 軸（axis 0）
+    sel = (ztop >= 0) & (np.arange(ny)[:, None] >= y_split)
+    yy, xx = np.where(sel)
+    if xx.size < 8:
+        return None
+    P = np.stack((xx, yy, ztop[yy, xx]), axis=1).astype(np.float64)
+    n = P.shape[0]
+
+    rng = np.random.default_rng(seed)
+    cos_min = np.cos(np.deg2rad(angle_max))
+    best_cnt, best_nv, best_d = -1, None, 0.0
+    for _ in range(n_iter):
+        i, j, k = rng.choice(n, 3, replace=False)
+        cr = np.cross(P[j] - P[i], P[k] - P[i])
+        ln = np.linalg.norm(cr)
+        if ln < 1e-6:
+            continue
+        nv = cr / ln
+        if nv[2] < 0:
+            nv = -nv
+        if nv[2] < cos_min:                               # 法線必須朝上
+            continue
+        d = float(nv @ P[i])
+        cnt = int(np.count_nonzero(np.abs(P @ nv - d) <= thresh))
+        if cnt > best_cnt:
+            best_cnt, best_nv, best_d = cnt, nv, d
+    if best_nv is None:
+        return None
+    # SVD 微調
+    inl = P[np.abs(P @ best_nv - best_d) <= thresh]
+    if inl.shape[0] < 3:
+        return None
+    mean = inl.mean(axis=0)
+    _, _, Vt = np.linalg.svd(inl - mean, full_matrices=False)
+    nv = Vt[2]
+    if nv[2] < 0:
+        nv = -nv
+    d = float(nv @ mean)
+    dist = np.abs(P @ nv - d)
+    inl = P[dist <= thresh]
+    u = np.cross(nv, [1.0, 0.0, 0.0])
+    u = u / np.linalg.norm(u)
+    v = np.cross(nv, u)
+    return {
+        'plane': (float(nv[0]), float(nv[1]), float(nv[2]), d),
+        'normal': (float(nv[0]), float(nv[1]), float(nv[2])),
+        'offset': d,
+        'tilt_deg': float(np.degrees(np.arccos(np.clip(nv[2], -1.0, 1.0)))),
+        'inlier_ratio': float(inl.shape[0] / n),
+        'n_points': int(n),
+        'n_inliers': int(inl.shape[0]),
+        'u': (float(u[0]), float(u[1]), float(u[2])),
+        'v': (float(v[0]), float(v[1]), float(v[2])),
+    }
+
 def azimuth_rotation(image, show_plt=False, save_plt=False, output_path=None):
 
     img = sitk.ReadImage(image, sitk.sitkUInt8)
@@ -55,6 +356,23 @@ def azimuth_rotation(image, show_plt=False, save_plt=False, output_path=None):
         plt.plot([x_center, cx], [y_min, cy], 'c-', lw=2, 
                 label=f'Angle with y-axis: {angle_deg:.1f}°')
 
+        # 中矢狀面：最佳鏡稱面 a·x + b·y + c·z = d（法線方向任意，不平行 YZ 面）
+        # 畫該平面在體積中央 z 切片上的截線
+        sym3 = best_symmetry_plane(arr_zyx)
+        a3, b3, c3, d3 = sym3['plane']
+        zc_ = (arr_zyx.shape[0] - 1) / 2.0
+        rhs_ = d3 - c3 * zc_                    # a*x + b*y = rhs
+        n2d = np.hypot(a3, b3)
+        x0_ = a3 * rhs_ / (n2d * n2d)
+        y0_ = b3 * rhs_ / (n2d * n2d)
+        u2_ = np.array([b3, -a3]) / n2d
+        tfg_ = (xs - x0_) * u2_[0] + (ys - y0_) * u2_[1]
+        L_ = float(np.abs(tfg_).max())
+        plt.plot([x0_ - L_ * u2_[0], x0_ + L_ * u2_[0]],
+                 [y0_ - L_ * u2_[1], y0_ + L_ * u2_[1]],
+                 color='magenta', lw=2, linestyle='--',
+                 label=f'Mirror plane: {a3:+.2f}x {b3:+.2f}y {c3:+.2f}z = {d3:.1f}')
+
         plt.legend()
         plt.title("Top point + centroid-directed line")
         plt.axis("off")
diff --git a/references/1-s2.0-S1529943008007213-main.pdf b/references/1-s2.0-S1529943008007213-main.pdf
new file mode 100644
index 0000000..5b4988d
Binary files /dev/null and b/references/1-s2.0-S1529943008007213-main.pdf differ
diff --git a/references/6_2021-0059.pdf b/references/6_2021-0059.pdf
new file mode 100644
index 0000000..28110cf
Binary files /dev/null and b/references/6_2021-0059.pdf differ
diff --git a/references/MIRU 2026/MIRU_poster_chou_CBT_v3.pdf b/references/MIRU 2026/MIRU_poster_chou_CBT_v3.pdf
new file mode 100644
index 0000000..51422d5
Binary files /dev/null and b/references/MIRU 2026/MIRU_poster_chou_CBT_v3.pdf differ
diff --git a/references/MIRU 2026/miru2026_CBT_v3.docx b/references/MIRU 2026/miru2026_CBT_v3.docx
new file mode 100644
index 0000000..6a2754c
Binary files /dev/null and b/references/MIRU 2026/miru2026_CBT_v3.docx differ
diff --git a/references/MIRU 2026/miru2026_CBT_v3.pdf b/references/MIRU 2026/miru2026_CBT_v3.pdf
new file mode 100644
index 0000000..5360160
Binary files /dev/null and b/references/MIRU 2026/miru2026_CBT_v3.pdf differ
diff --git a/references/asj-11-817.pdf b/references/asj-11-817.pdf
new file mode 100644
index 0000000..64d4186
Binary files /dev/null and b/references/asj-11-817.pdf differ
diff --git a/visualization/res_plot_3d.py b/visualization/res_plot_3d.py
index fe4cd9d..3ab072d 100644
--- a/visualization/res_plot_3d.py
+++ b/visualization/res_plot_3d.py
@@ -1,6 +1,9 @@
 import torch
 import numpy as np
 import matplotlib.pyplot as plt
+from matplotlib.colors import to_rgba
+from matplotlib.lines import Line2D
+from mpl_toolkits.mplot3d.art3d import Poly3DCollection
 import os
 from datetime import datetime
 import csv
@@ -8,9 +11,20 @@ import csv
 from core.cylinder import generate_cylinder_n_torch, generate_cylinder_o_torch, snap_to_discrete_values
 from core.intersection import center_line_intersections_torch
 from core.scoring import cl_score_torch, compute_overlap_ratio_from_cylinder_mask, cl_score_torch_xfr
-from imaging.orientation import azimuth_rotation, analyze_vertebral_tilt_contour
+from imaging.orientation import (azimuth_rotation, analyze_vertebral_tilt_contour,
+                                  best_symmetry_plane, best_upper_endplate_plane,
+                                  segment_spinous_process)
 from utils.helpers import save_with_unique_name
 
+# Volume absorption 渲染（Beer-Lambert）：每 voxel 不透明度 = 1 - exp(-mu * voxel_width)
+# 骨頭核心厚度達 70-90 voxel，沿視線堆疊會使任何 per-voxel alpha 累積成不透明。
+# 因此以「抽稀 (SUBSAMPLE) 降低堆疊數量」+「低 mu 控制每點吸收」兩項共同調出淡薄 X-ray 陰影，
+# 同時保留皮質 / 鬆質的吸收入射差異（mu 比值）。
+BONE_MU_CORTICAL = 0.02    # 1/mm  → 每 voxel = 1-exp(-0.02*0.5) ~ 0.010
+BONE_MU_TRABECULAR = 0.005 # 1/mm  → 每 voxel = 1-exp(-0.005*0.5) ~ 0.0025
+BONE_MARKER_SIZE = 3.0     # 骨骼散點點面積 (pt^2)；略大以補償抽稀後的顆粒感
+BONE_SUBSAMPLE = 1         # 每 10 個骨 voxel 畫 1 個，降低堆疊不透明度（1=全畫）
+
 def set_axes_equal_3d(ax):
     """
     Make axes of 3D plot have equal scale so that spheres appear as spheres,
@@ -165,55 +179,171 @@ def res_plt_2_torch(
     z_cyl_r1, y_cyl_r1, x_cyl_r1 = np.where(cyl_r_cpu == 1)
     z_cyl_r2, y_cyl_r2, x_cyl_r2 = np.where(cyl_ro_cpu == 1)
 
-    z_img, y_img, x_img = np.where(spine_cpu == 1)
+    # 骨頭 voxel 依「體積吸收」分成兩組：皮質（高不透明度）與鬆質（低不透明度）
+    cortical_cpu = cortical_tensor.cpu().numpy()
+    voxel_mm = float(spacing[0])
+    alpha_cortical = 1.0 - np.exp(-BONE_MU_CORTICAL * voxel_mm)
+    alpha_trabecular = 1.0 - np.exp(-BONE_MU_TRABECULAR * voxel_mm)
+    z_corti, y_corti, x_corti = np.where((spine_cpu == 1) & (cortical_cpu == 1))
+    z_trab, y_trab, x_trab = np.where((spine_cpu == 1) & (cortical_cpu == 0))
+
+    # 中矢狀面：骨頭的最佳鏡稱面，一般平面 a·x + b·y + c·z = d（法線方向任意）
+    sym = best_symmetry_plane(spine_cpu)
+
+    # 棘突：鏡稱面中線帶（|s|<=w）且在 AP 谷底之後側的骨 voxel，換不同顏色標示
+    sp_mask, sp_th, sp_info = segment_spinous_process(spine_cpu, sym)
+    if sp_mask is not None and sp_mask.any():
+        sp_corti = sp_mask[z_corti, y_corti, x_corti]
+        sp_trab = sp_mask[z_trab, y_trab, x_trab]
+        sp_n_bone = max(int(spine_cpu.sum()), 1)
+        print(f"[SPINOUS] n={sp_info['n_sp']} "
+              f"({100.0 * sp_info['n_sp'] / sp_n_bone:.1f}% of bone) "
+              f"band=+/-{sp_info['band_w']:.1f} voxel  AP>={sp_info['ap_thresh']:.1f} "
+              f"mode={sp_info['mode']}")
+    else:
+        sp_corti = sp_trab = None
+
+    # X-ray 外觀：骨頭合成一個半透明體積吸收點雲（下方）；
+    # 螺絲（中心線 + 圓柱 + 入口軌跡延长）合成一個全不透明點雲，永遠畫在骨頭之上
+    def _rgba_block(n, color, a):
+        arr = np.empty((n, 4))
+        arr[:] = to_rgba(color)
+        arr[:, 3] = a
+        return arr
+
+    x_bone = np.concatenate([x_corti, x_trab])
+    y_bone = np.concatenate([y_corti, y_trab])
+    z_bone = np.concatenate([z_corti, z_trab])
+    bone_rgba = np.concatenate([
+        _rgba_block(len(x_corti), 'lightblue', float(alpha_cortical)),
+        _rgba_block(len(x_trab), 'lightblue', float(alpha_trabecular)),
+    ])
+    bone_size = np.full(len(x_bone), BONE_MARKER_SIZE)
+
+    # 抽稀：降低堆疊不透明度以呈現淡薄 X-ray 陰影
+    if BONE_SUBSAMPLE > 1:
+        x_bone = x_bone[::BONE_SUBSAMPLE]
+        y_bone = y_bone[::BONE_SUBSAMPLE]
+        z_bone = z_bone[::BONE_SUBSAMPLE]
+        bone_rgba = bone_rgba[::BONE_SUBSAMPLE]
+        bone_size = bone_size[::BONE_SUBSAMPLE]
+    if sp_corti is not None:
+        sp_flag = np.concatenate([sp_corti, sp_trab])
+        if BONE_SUBSAMPLE > 1:
+            sp_flag = sp_flag[::BONE_SUBSAMPLE]
+        bone_rgba[sp_flag] = to_rgba('purple', 0.95)
+
+    x_screw = np.concatenate([x_lin1, x_lin2, x_cyl_l1, x_cyl_l2, x_cyl_r1, x_cyl_r2])
+    y_screw = np.concatenate([y_lin1, y_lin2, y_cyl_l1, y_cyl_l2, y_cyl_r1, y_cyl_r2])
+    z_screw = np.concatenate([z_lin1, z_lin2, z_cyl_l1, z_cyl_l2, z_cyl_r1, z_cyl_r2])
+
+    _a, _b, _c, _d = sym['plane']
+    _n = np.array([_a, _b, _c])
+    _u = np.array(sym['u'])
+    _v = np.array(sym['v'])
+    _p0 = _d * _n                                # 平面上最接近原點的點
+    xyz_bone = np.stack([x_bone - _p0[0], y_bone - _p0[1], z_bone - _p0[2]], axis=1)
+    _pu = xyz_bone @ _u
+    _pv = xyz_bone @ _v
+    _U_, _V_ = np.meshgrid(np.linspace(_pu.min(), _pu.max(), 8),
+                           np.linspace(_pv.min(), _pv.max(), 8))
+    _Xp = _p0[0] + _U_ * _u[0] + _V_ * _v[0]
+    _Yp = _p0[1] + _U_ * _u[1] + _V_ * _v[1]
+    _Zp = _p0[2] + _U_ * _u[2] + _V_ * _v[2]
+
+    # 上終板平面：RANSAC 擬合骨頭頂面（前側）的最佳 a·x + b·y + c·z = d
+    symp = best_upper_endplate_plane(spine_cpu)
+    _EX = _EY = _EZ = None
+    if symp is not None:
+        _ea, _eb, _ec, _ed = symp['plane']
+        _en = np.array([_ea, _eb, _ec])
+        _eu = np.array(symp['u'])
+        _ev = np.array(symp['v'])
+        _ep0 = _ed * _en
+        xz_ep = np.stack([x_bone - _ep0[0], y_bone - _ep0[1], z_bone - _ep0[2]], axis=1)
+        _pu_ep = xz_ep @ _eu
+        _pv_ep = xz_ep @ _ev
+        _EU, _EV = np.meshgrid(np.linspace(_pu_ep.min(), _pu_ep.max(), 8),
+                               np.linspace(_pv_ep.min(), _pv_ep.max(), 8))
+        _EX = _ep0[0] + _EU * _eu[0] + _EV * _ev[0]
+        _EY = _ep0[1] + _EU * _eu[1] + _EV * _ev[1]
+        _EZ = _ep0[2] + _EU * _eu[2] + _EV * _ev[2]
+    screw_rgba = np.concatenate([
+        _rgba_block(len(x_lin1), 'r', 1.0),
+        _rgba_block(len(x_lin2), 'r', 1.0),
+        _rgba_block(len(x_cyl_l1), 'darkcyan', 1.0),
+        _rgba_block(len(x_cyl_l2), 'pink', 1.0),
+        _rgba_block(len(x_cyl_r1), 'blue', 1.0),
+        _rgba_block(len(x_cyl_r2), 'pink', 1.0),
+    ])
+    screw_size = np.concatenate([
+        np.full(len(x_lin1), 3), np.full(len(x_lin2), 3),
+        np.full(len(x_cyl_l1), 36), np.full(len(x_cyl_l2), 36),
+        np.full(len(x_cyl_r1), 36), np.full(len(x_cyl_r2), 36),
+    ])
 
     fig = plt.figure(figsize=(12, 12))
 
+    # 圖例色塊提高到可讀不透明度（實際渲染仍用真實吸收 alpha）
+    _leg_alpha_c = max(float(alpha_cortical), 0.35)
+    _leg_alpha_t = max(float(alpha_trabecular), 0.2)
+    legend_handles = [
+        Line2D([], [], marker='o', ls='', ms=5, color=to_rgba('lightblue', _leg_alpha_c), label='Spine (cortical)'),
+        Line2D([], [], marker='o', ls='', ms=5, color=to_rgba('lightblue', _leg_alpha_t), label='Spine (trabecular)'),
+        Line2D([], [], marker='o', ls='', ms=2, color='r', label='Centerline'),
+        Line2D([], [], marker='o', ls='', ms=6, color='darkcyan', label='Cylinder(L)'),
+        Line2D([], [], marker='o', ls='', ms=6, color='blue', label='Cylinder(R)'),
+        Line2D([], [], marker='o', ls='', ms=6, color='pink', label='Entry track (outer)'),
+Line2D([], [], color='orange', lw=2, alpha=0.6,
+        label=f"Mirror plane {sym['plane'][0]:+.2f}x {sym['plane'][1]:+.2f}y {sym['plane'][2]:+.2f}z = {sym['plane'][3]:.1f}"),
+    ]
+    if sp_corti is not None:
+        legend_handles.append(Line2D([], [], marker='o', ls='', ms=6, color='purple',
+                              label=f"Spinous process (mirror-plane midline, {sp_info['n_sp']} vox)"))
+    if symp is not None:
+        legend_handles.append(Line2D([], [], color='green', lw=2, alpha=0.7,
+                              label=f"Upper endplate plane tilt {symp['tilt_deg']:.1f} deg"))
+
+    def _fill_ax(ax):
+        # X-ray 外觀：關閉 mplot3d 依深度自動排序 zorder（否則半透明骨頭會被重繪到
+        # 螺絲上方）；改為固定分層：吸收骨頭 zorder=5，全不透明螺絲 zorder=10
+        ax.computed_zorder = False
+        sc_bone = ax.scatter(x_bone, y_bone, z_bone, c=bone_rgba, s=bone_size, marker='o')
+        sc_bone.set_zorder(5)
+        sc_screw = ax.scatter(x_screw, y_screw, z_screw, c=screw_rgba, s=screw_size, marker='o')
+        sc_screw.set_zorder(10)
+        # 中矢狀面（理論左右對稱切分面）：半透明橘色平面 x = x_mid
+        # 平面邊緣畫橘色線，讓 axial / 正視（側看時）也能清楚看到切分線
+        plane = ax.plot_surface(_Xp, _Yp, _Zp, color='orange', alpha=0.30,
+                                linewidth=1.0, edgecolor='orange', rstride=1, cstride=1)
+        plane.set_zorder(8)
+        # 上終板平面：半透明綠色平面（邊緣綠線）
+        if _EX is not None:
+            ep = ax.plot_surface(_EX, _EY, _EZ, color='green', alpha=0.35,
+                                 linewidth=1.0, edgecolor='green', rstride=1, cstride=1)
+            ep.set_zorder(7)
+
     ax1 = fig.add_subplot(221, projection='3d')
-    ax1.scatter(x_lin1, y_lin1, z_lin1, c='r', marker='o', s=1)
-    ax1.scatter(x_lin2, y_lin2, z_lin2, c='r', marker='o', s=1)
-    ax1.scatter(x_cyl_l1, y_cyl_l1, z_cyl_l1, c='darkcyan', marker='o', label='Cylinder(L)')
-    ax1.scatter(x_cyl_l2, y_cyl_l2, z_cyl_l2, c='pink', marker='o')
-    ax1.scatter(x_cyl_r1, y_cyl_r1, z_cyl_r1, c='blue', marker='o', label='Cylinder(R)')
-    ax1.scatter(x_cyl_r2, y_cyl_r2, z_cyl_r2, c='pink', marker='o')
-    ax1.scatter(x_img, y_img, z_img, c='lightblue', marker='+', alpha=0.04, label='Spine')
+    _fill_ax(ax1)
     ax1.set_xlabel('X-axis'); ax1.set_ylabel('Y-axis'); ax1.set_zlabel('Z-axis')
     set_axes_equal_3d(ax1)
 
     ax2 = fig.add_subplot(222, projection='3d')
     ax2.view_init(elev=90, azim=-90, roll=0)
-    ax2.scatter(x_lin1, y_lin1, z_lin1, c='r', marker='o', s=1)
-    ax2.scatter(x_lin2, y_lin2, z_lin2, c='r', marker='o', s=1)
-    ax2.scatter(x_cyl_l1, y_cyl_l1, z_cyl_l1, c='darkcyan', marker='o', label='Cylinder(L)')
-    ax2.scatter(x_cyl_l2, y_cyl_l2, z_cyl_l2, c='pink', marker='o')
-    ax2.scatter(x_cyl_r1, y_cyl_r1, z_cyl_r1, c='blue', marker='o', label='Cylinder(R)')
-    ax2.scatter(x_cyl_r2, y_cyl_r2, z_cyl_r2, c='pink', marker='o')
-    ax2.scatter(x_img, y_img, z_img, c='lightblue', marker='+', alpha=0.04, label='Spine')
+    _fill_ax(ax2)
     ax2.set_xlabel('X-axis'); ax2.set_ylabel('Y-axis'); ax2.set_zlabel('Z-axis')
     set_axes_equal_3d(ax2)
-    ax2.legend()
+    ax2.legend(handles=legend_handles)
 
     ax3 = fig.add_subplot(223, projection='3d')
     ax3.view_init(elev=0, azim=90, roll=0)
-    ax3.scatter(x_lin1, y_lin1, z_lin1, c='r', marker='o', s=1)
-    ax3.scatter(x_lin2, y_lin2, z_lin2, c='r', marker='o', s=1)
-    ax3.scatter(x_cyl_l1, y_cyl_l1, z_cyl_l1, c='darkcyan', marker='o', label='Cylinder(L)')
-    ax3.scatter(x_cyl_l2, y_cyl_l2, z_cyl_l2, c='pink', marker='o')
-    ax3.scatter(x_cyl_r1, y_cyl_r1, z_cyl_r1, c='blue', marker='o', label='Cylinder(R)')
-    ax3.scatter(x_cyl_r2, y_cyl_r2, z_cyl_r2, c='pink', marker='o')
-    ax3.scatter(x_img, y_img, z_img, c='lightblue', marker='+', alpha=0.04, label='Spine')
+    _fill_ax(ax3)
     ax3.set_xlabel('X-axis'); ax3.set_ylabel('Y-axis'); ax3.set_zlabel('Z-axis')
     set_axes_equal_3d(ax3)
 
     ax4 = fig.add_subplot(224, projection='3d')
     ax4.view_init(elev=0, azim=0, roll=0)
-    ax4.scatter(x_lin1, y_lin1, z_lin1, c='r', marker='o', s=1)
-    ax4.scatter(x_lin2, y_lin2, z_lin2, c='r', marker='o', s=1)
-    ax4.scatter(x_cyl_l1, y_cyl_l1, z_cyl_l1, c='darkcyan', marker='o', label='Cylinder(L)')
-    ax4.scatter(x_cyl_l2, y_cyl_l2, z_cyl_l2, c='pink', marker='o')
-    ax4.scatter(x_cyl_r1, y_cyl_r1, z_cyl_r1, c='blue', marker='o', label='Cylinder(R)')
-    ax4.scatter(x_cyl_r2, y_cyl_r2, z_cyl_r2, c='pink', marker='o')
-    ax4.scatter(x_img, y_img, z_img, c='lightblue', marker='+', alpha=0.04, label='Spine')
+    _fill_ax(ax4)
     ax4.set_xlabel('X-axis'); ax4.set_ylabel('Y-axis'); ax4.set_zlabel('Z-axis')
     set_axes_equal_3d(ax4)
     
@@ -225,12 +355,12 @@ def res_plt_2_torch(
     overlap_b_l = ((spine_tensor == 1) & (cyl_l == 1)).sum().item()
     overlap_b_r = ((spine_tensor == 1) & (cyl_r == 1)).sum().item()
 
-    overlap_cortical_l = (overlap_l / cyl_points_l) * 100
-    overlap_cortical_r = (overlap_r / cyl_points_r) * 100
-    overlap_vertebral_l = (overlap_b_l / cyl_points_l) * 100
-    overlap_vertebral_r = (overlap_b_r / cyl_points_r) * 100
-    cb_ratio_l = overlap_cortical_l/overlap_vertebral_l
-    cb_ratio_r = overlap_cortical_r/overlap_vertebral_r
+    overlap_cortical_l = (overlap_l / cyl_points_l) * 100 if cyl_points_l else 0.0
+    overlap_cortical_r = (overlap_r / cyl_points_r) * 100 if cyl_points_r else 0.0
+    overlap_vertebral_l = (overlap_b_l / cyl_points_l) * 100 if cyl_points_l else 0.0
+    overlap_vertebral_r = (overlap_b_r / cyl_points_r) * 100 if cyl_points_r else 0.0
+    cb_ratio_l = overlap_cortical_l/overlap_vertebral_l if overlap_vertebral_l else 0.0
+    cb_ratio_r = overlap_cortical_r/overlap_vertebral_r if overlap_vertebral_r else 0.0
     user_altitude_l = 90 - best_position_l[4] - alt
     user_altitude_r = 90 - best_position_r[4] - alt
     user_azimuth_l = 90 - best_position_l[3] - azi
diff --git a/xfr_debug.py b/xfr_debug.py
index eaee490..581dfe4 100644
--- a/xfr_debug.py
+++ b/xfr_debug.py
@@ -1,4 +1,10 @@
 import os
+import re
+import sys
+import time
+import queue as queue_module
+import subprocess
+import multiprocessing as mp
 
 import SimpleITK as sitk
 import torch
@@ -15,6 +21,108 @@ azimuth_rotation_dir = '/mnt/1248/open2/cyrou/azimuth_rotation'
 tilt_contour_dir = '/mnt/1248/open2/cyrou/tilt_contour'
 Output_dir = '/mnt/1248/open/cyrou/Output'
 
+LEVELS = ('L1', 'L2', 'L3', 'L4', 'L5')
+
+LOG_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'logs')
+
+
+# 目前任務標記（每個 worker 流程各自一份），讓交錯的 log 可以歸屬到 (volume, level, side)
+_TASK_TAG = {'vid': None, 'level': None, 'side': None}
+
+_NP_WRAP_RE = re.compile(r'\bnp\.[A-Za-z_][A-Za-z0-9_]*\(([^()]*)\)')
+
+
+def set_task_tag(volume_id, level):
+    _TASK_TAG['vid'] = volume_id
+    _TASK_TAG['level'] = level
+    _TASK_TAG['side'] = None
+
+
+class _Tee:
+    """同時輸出到 console 與 log 檔，逐行加上如 [0001 L1 LEFT] 的標記。
+    volume 取 id 末段（1.3.6.1.4.1.9328.50.4.0001 -> 0001）；
+    side 由 [LEFT]/[RIGHT]/左側/右側 段落標記判定並沿用給後續行；
+    最終結果段屬於整椎，重置沿用值，其中的 Left/Right 摘要行只標該行本身；
+    np.float64(...) 之類包裹會解包成裸數值。"""
+
+    def __init__(self, console, log_fh):
+        self.console = console
+        self.log_fh = log_fh
+        self.buf = ''
+
+    @staticmethod
+    def _marker_side(line):
+        """回傳 (side, persist)：[LEFT]/[RIGHT]/左側/右側 是段落標記（persist=True），
+        Left/Right 摘要行只標該行（persist=False）"""
+        s = line.lstrip()
+        if s.startswith('[LEFT]') or '左側' in s:
+            return 'LEFT', True
+        if s.startswith('[RIGHT]') or '右側' in s:
+            return 'RIGHT', True
+        if s.startswith('Left '):
+            return 'LEFT', False
+        if s.startswith('Right '):
+            return 'RIGHT', False
+        return None, False
+
+    def _prefix(self, side):
+        if not _TASK_TAG['vid'] or not _TASK_TAG['level']:
+            return ''
+        vol = _TASK_TAG['vid'].rsplit('.', 1)[-1]
+        tag = f'{vol} {_TASK_TAG["level"]}'
+        if side:
+            tag += f' {side}'
+        return f'[{tag}] '
+
+    def _emit(self, line):
+        line = _NP_WRAP_RE.sub(r'\1', line)
+        if '最終結果' in line:
+            _TASK_TAG['side'] = None
+        side, persist = self._marker_side(line)
+        if persist:
+            _TASK_TAG['side'] = side
+        eff_side = side if side is not None else _TASK_TAG['side']
+        prefix = self._prefix(eff_side) if line.strip() else ''
+        if prefix:
+            # 前綴已含 side 時，去掉行首重複的 [LEFT] / [RIGHT] 標記
+            marker = f'[{eff_side}] '
+            if line.startswith(marker):
+                line = line[len(marker):]
+        self.console.write(prefix + line + '\n')
+        self.log_fh.write(prefix + line + '\n')
+
+    def write(self, data):
+        if not data:
+            return
+        self.buf += data
+        while True:
+            idx_n = self.buf.find('\n')
+            idx_r = self.buf.find('\r')
+            candidates = [i for i in (idx_n, idx_r) if i != -1]
+            if not candidates:
+                break
+            idx = min(candidates)
+            self._emit(self.buf[:idx])
+            self.buf = self.buf[idx + 1:]
+
+    def flush(self):
+        if self.buf:
+            self._emit(self.buf)
+            self.buf = ''
+        self.console.flush()
+        self.log_fh.flush()
+
+    def isatty(self):
+        return False
+
+
+def setup_tee(log_path):
+    """把這個流程的 stdout/stderr 同時寫到 log_path（append、line-buffered）"""
+    log_fh = open(log_path, 'a', buffering=1)
+    sys.stdout = _Tee(sys.stdout, log_fh)
+    sys.stderr = _Tee(sys.stderr, log_fh)
+
+
 def get_device(gpu_id=None):
     if torch.cuda.is_available():
         if gpu_id is None:
@@ -22,7 +130,7 @@ def get_device(gpu_id=None):
             gpu_id = 0
             for i in range(torch.cuda.device_count()):
                 free_mem, _ = torch.cuda.mem_get_info(i)
-                # print(f'GPU {i}: {torch.cuda.get_device_name(i)} {free_mem}')
+                # print(f'GPU {gpu_id}: {torch.cuda.get_device_name(gpu_id)} {free_mem}')
                 if free_mem > max_free:
                     max_free = free_mem
                     gpu_id = i
@@ -52,14 +160,15 @@ def debug_orientation(volume_id, level):
     # print(f'Azimuth: {azi}, Alt: {alt}')
     print(f'Alt: {alt}')
 
-def debug_pso(volume_id, level):
+def debug_pso(volume_id, level, device=None):
     # ====== PSO ======
     swarm_size = 100
     max_iter = 100
 
     # ====== DEVICE ======
-    device = get_device()  
-
+    if device is None:
+        device = get_device()
+  
     # ====== OTHER ======
     spacing = [0.5, 0.5, 0.5]
     CBT = True
@@ -76,9 +185,15 @@ def debug_pso(volume_id, level):
 
     cortical_array = sitk.GetArrayFromImage(cortical_image)
     binary_array = sitk.GetArrayFromImage(binary_image)
-    roi_array = sitk.GetArrayFromImage(roi_image)  
+    roi_array = sitk.GetArrayFromImage(roi_image)
     image_shape = binary_array.shape
 
+    # 資料層級的快速失敗（例：…9328.50.4.0653 L5 是 1-voxel 寬的退化體積）
+    if binary_array.sum() == 0:
+        raise ValueError(f'{volume_id} {level}: empty bone mask')
+    if min(image_shape) < 8:
+        raise ValueError(f'{volume_id} {level}: degenerate volume shape {image_shape}')
+
     cortical_tensor = torch.tensor(cortical_array, device=device)
     binary_tensor = torch.tensor(binary_array, device=device)
 
@@ -93,7 +208,7 @@ def debug_pso(volume_id, level):
         grid_=grid,
         use_tip_penalty=False 
     )    
-
+    
     best_l, loss_l, best_r, loss_r, total_time = run_pso_torch_xfr(
         label_str=f'{volume_id} {level}',
         image1_path=cortical_path,
@@ -112,6 +227,57 @@ def debug_pso(volume_id, level):
     # exit()
 
 
+def list_gpu_ids():
+    """透過 nvidia-smi 取得 GPU ID（主流程不初始化 CUDA，避免污染 fork/spawn 子流程）"""
+    try:
+        out = subprocess.check_output(
+            ['nvidia-smi', '--query-gpu=index', '--format=csv,noheader'],
+            text=True,
+        )
+        return [int(line.strip()) for line in out.splitlines() if line.strip()]
+    except Exception:
+        return []
+
+
+def gpu_worker(gpu_id, log_path, task_queue, result_queue):
+    """每張 GPU 一個工作流程：先鎖死該 GPU，再從共享隊列領 (volume, level) 任務"""
+    # worker 是獨立流程，要自己把輸出 tee 到 log 檔
+    setup_tee(log_path)
+    # 必須在任何 torch.cuda 呼叫前設定
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_id)
+    torch.cuda.set_device(0)
+    device = torch.device('cuda:0')
+    print(f'=== [GPU {gpu_id}] worker started ===', flush=True)
+
+    while True:
+        # 注意：multiprocessing.Queue 沒有 task_done()（只有 queue.Queue 有），別加回來
+        item = task_queue.get()
+        if item is None:
+            break
+        volume_id, level = item
+        set_task_tag(volume_id, level)
+        try:
+            debug_pso(volume_id, level, device)
+            result_queue.put(('task', gpu_id, volume_id, level, True, ''))
+        except Exception as e:
+            print(f'[GPU {gpu_id}] Error in {volume_id} {level}: {e}', flush=True)
+            result_queue.put(('task', gpu_id, volume_id, level, False, str(e)))
+
+    result_queue.put(('done', gpu_id))
+    print(f'=== [GPU {gpu_id}] worker finished ===', flush=True)
+
+
+def _run_sequential(tasks):
+    """沒有（或只有一張）GPU 時的回退：單流程串行"""
+    device = get_device()
+    for volume_id, level in tasks:
+        set_task_tag(volume_id, level)
+        try:
+            debug_pso(volume_id, level, device)
+        except Exception as e:
+            print(f'Error in {volume_id} {level}: {e}')
+
+
 def main():
     # level = 'L1'
 
@@ -121,38 +287,97 @@ def main():
     # process_volume(volume_id, level)
     # exit()
 
+    # 要處理的 volume 數（並行模式下是「最多嘗試的 volume 數」）
     MAX_SUCCESSFUL_VOLUMES = 100
-    number_of_successful_volumes = 0
+    MAX_SUCCESSFUL_VOLUMES = 1
 
-    # for volume_id in (
-    #     '1.3.6.1.4.1.9328.50.4.0001', 
-    #     '1.3.6.1.4.1.9328.50.4.0002', 
-    #     '1.3.6.1.4.1.9328.50.4.0003', 
-    #     '1.3.6.1.4.1.9328.50.4.0004',
-    #     '1.3.6.1.4.1.9328.50.4.0005',
-    #     # '1.3.6.1.4.1.9328.50.4.0006',
-    # ):
+    # log 檔（console 與檔案同時輸出；各 GPU worker 也會 append 進同一個檔）
+    os.makedirs(LOG_DIR, exist_ok=True)
+    log_path = os.path.join(LOG_DIR, f'xfr_debug_{time.strftime("%Y%m%d_%H%M%S")}.log')
+    setup_tee(log_path)
+    print(f'Log file: {log_path}', flush=True)
+    print(f'Command: {sys.executable} {" ".join(sys.argv)}', flush=True)
+    print(f'Working directory: {os.getcwd()}', flush=True)
 
-    for volume_id in sorted(os.listdir(standardized_dir)):
-        # debug_orientation(volume_id, level)
-        error_flag = False
-        for level in ('L1', 'L2', 'L3', 'L4', 'L5'):
-        # for level in ('L5',):
-            # debug_orientation(volume_id, level)
-            try:
-                debug_pso(volume_id, level)
-            except Exception as e:
-                print(f'Error in {volume_id} {level}: {e}')
-                error_flag = True
-                continue
-        if not error_flag:
-            number_of_successful_volumes += 1
+    volumes = [d for d in sorted(os.listdir(standardized_dir))
+               if os.path.isdir(os.path.join(standardized_dir, d))]
+    volumes = volumes[:MAX_SUCCESSFUL_VOLUMES]
 
-        if number_of_successful_volumes >= MAX_SUCCESSFUL_VOLUMES:
+    tasks = [(vid, level) for vid in volumes for level in LEVELS]
+    print(f'Total {len(volumes)} volumes / {len(tasks)} (volume, level) tasks', flush=True)
+
+    gpu_ids = list_gpu_ids()
+
+    if len(gpu_ids) <= 1:
+        print(f'Only {len(gpu_ids)} GPU(s) available, running sequentially', flush=True)
+        _run_sequential(tasks)
+        return
+
+    print(f'Found {len(gpu_ids)} GPUs: {gpu_ids}, starting {len(gpu_ids)} workers (one per GPU)', flush=True)
+
+    ctx = mp.get_context('spawn')
+    task_queue = ctx.Queue()
+    result_queue = ctx.Queue()
+    for t in tasks:
+        task_queue.put(t)
+    for _ in gpu_ids:
+        task_queue.put(None)  # 每個 worker 一個結束哨兵
+
+    procs = [ctx.Process(target=gpu_worker, args=(g, log_path, task_queue, result_queue), name=f'cbt-gpu-{g}')
+             for g in gpu_ids]
+    for p in procs:
+        p.start()
+
+    start_time = time.time()
+    results = []
+    finished = 0
+    while finished < len(gpu_ids):
+        if not any(p.is_alive() for p in procs):
+            print('Warning: a worker exited early; reaping remaining tasks...', flush=True)
             break
+        try:
+            msg = result_queue.get(timeout=5)
+        except queue_module.Empty:
+            continue
+        if msg[0] == 'done':
+            finished += 1
+        else:
+            results.append(msg)
 
-        # exit()
+    # 抽乾剩下排進來的結果
+    while True:
+        try:
+            msg = result_queue.get(timeout=1)
+        except queue_module.Empty:
+            break
+        if msg[0] == 'task':
+            results.append(msg)
 
+    for p in procs:
+        p.join(timeout=60)
+
+    total_time = time.time() - start_time
+    ok = [r for r in results if r[4]]
+    fail = [r for r in results if not r[4]]
+    per_volume = {}
+    for _, _, vid, level, success, _ in results:
+        per_volume.setdefault(vid, set()).add(success)
+    missing = len(tasks) - len(results)
+
+    # 一個 volume 算「成功」必須它的(level)全部執行過且全部成功
+    n_success_volumes = sum(1 for vid in volumes
+                            if per_volume.get(vid, set()) == {True})
+
+    print('=' * 60)
+    print(f'Finished in {total_time / 60:.1f} min | '
+          f'tasks {len(results)}/{len(tasks)} (ok {len(ok)} / failed {len(fail)} / not-run {missing})')
+    if missing:
+        print(f'Warning: {missing} task(s) were never executed (worker crash?)')
+    print(f'Successful volumes (all levels OK): {n_success_volumes}/{len(volumes)}')
+    if fail:
+        print('Failed tasks:')
+        for _, g, vid, level, _, err in fail:
+            print(f'  [GPU {g}] {vid} {level}: {err}')
 
 
 if __name__ == '__main__':