Add official VoxelMorph CT registration pipeline

metrics.py

@@ -0,0 +1,123 @@
+"""配准质量评估工具。
+
+这里刻意只依赖 NumPy，便于在训练、推理和 Streamlit 三处复用。
+"""
+
+from __future__ import annotations
+
+from typing import Dict, Iterable, Tuple
+
+import numpy as np
+
+
+def crop_to_common_shape(*arrays: np.ndarray) -> Tuple[np.ndarray, ...]:
+    """把多个 3D/4D 数组中心裁剪到共同尺寸，避免形状不一致导致评估失败。"""
+
+    if not arrays:
+        return tuple()
+
+    spatial_shapes = [arr.shape[:3] for arr in arrays]
+    common = tuple(min(shape[axis] for shape in spatial_shapes) for axis in range(3))
+    cropped = []
+
+    for arr in arrays:
+        slices = []
+        for axis, target in enumerate(common):
+            start = max((arr.shape[axis] - target) // 2, 0)
+            slices.append(slice(start, start + target))
+        if arr.ndim > 3:
+            slices.append(slice(None))
+        cropped.append(arr[tuple(slices)])
+
+    return tuple(cropped)
+
+
+def global_ncc(a: np.ndarray, b: np.ndarray, eps: float = 1e-8) -> float:
+    """计算全局 NCC。训练时使用局部 NCC，这里用于快速量化展示。"""
+
+    a = np.asarray(a, dtype=np.float32)
+    b = np.asarray(b, dtype=np.float32)
+    a = a - float(np.mean(a))
+    b = b - float(np.mean(b))
+    denom = float(np.sqrt(np.sum(a * a) * np.sum(b * b)) + eps)
+    return float(np.sum(a * b) / denom)
+
+
+def mse(a: np.ndarray, b: np.ndarray) -> float:
+    diff = np.asarray(a, dtype=np.float32) - np.asarray(b, dtype=np.float32)
+    return float(np.mean(diff * diff))
+
+
+def mae(a: np.ndarray, b: np.ndarray) -> float:
+    diff = np.abs(np.asarray(a, dtype=np.float32) - np.asarray(b, dtype=np.float32))
+    return float(np.mean(diff))
+
+
+def registration_metrics(
+    fixed: np.ndarray,
+    moving: np.ndarray,
+    warped: np.ndarray,
+) -> Dict[str, float]:
+    """输出配准前后可比较的常用指标。"""
+
+    fixed, moving, warped = crop_to_common_shape(fixed, moving, warped)
+    before_mse = mse(fixed, moving)
+    after_mse = mse(fixed, warped)
+    before_mae = mae(fixed, moving)
+    after_mae = mae(fixed, warped)
+    before_ncc = global_ncc(fixed, moving)
+    after_ncc = global_ncc(fixed, warped)
+
+    return {
+        "before_mse": before_mse,
+        "after_mse": after_mse,
+        "before_mae": before_mae,
+        "after_mae": after_mae,
+        "before_ncc": before_ncc,
+        "after_ncc": after_ncc,
+        "mse_improvement": before_mse - after_mse,
+        "mae_improvement": before_mae - after_mae,
+        "ncc_improvement": after_ncc - before_ncc,
+    }
+
+
+def slice_metric_curve(
+    fixed: np.ndarray,
+    moving: np.ndarray,
+    warped: np.ndarray,
+    axis: int = 2,
+) -> Dict[str, Iterable[float]]:
+    """逐切片计算 MSE，适合生成“配准前后误差曲线”。"""
+
+    fixed, moving, warped = crop_to_common_shape(fixed, moving, warped)
+    before = []
+    after = []
+
+    for index in range(fixed.shape[axis]):
+        selector = [slice(None)] * 3
+        selector[axis] = index
+        selector = tuple(selector)
+        before.append(mse(fixed[selector], moving[selector]))
+        after.append(mse(fixed[selector], warped[selector]))
+
+    return {
+        "slice_index": list(range(fixed.shape[axis])),
+        "before_mse": before,
+        "after_mse": after,
+    }
+
+
+def ddf_summary(ddf_xyz: np.ndarray) -> Dict[str, float]:
+    """统计形变场向量大小。输入应为 X/Y/Z/3，单位可为 voxel 或 mm。"""
+
+    if ddf_xyz.ndim != 4 or ddf_xyz.shape[-1] != 3:
+        raise ValueError("DDF 必须是形状为 (X, Y, Z, 3) 的 4D 数组。")
+
+    mag = np.linalg.norm(ddf_xyz.astype(np.float32), axis=-1)
+    return {
+        "ddf_mean": float(np.mean(mag)),
+        "ddf_std": float(np.std(mag)),
+        "ddf_p95": float(np.percentile(mag, 95)),
+        "ddf_max": float(np.max(mag)),
+    }
+