Td/46accuracy_roi

.gitignore

@@ -0,0 +1,7 @@
+*wandb*
+*totalspineseg*
+*model*
+*mil*
+*inference*
+*__pycache__*
+*csv*

metrics/ROIaccuracy.py

@@ -0,0 +1,130 @@
+import os
+import argparse
+import numpy as np
+import pandas as pd
+import pydicom
+
+def get_axis_labels(row_cos, col_cos, normal):
+    """
+    Determine the anatomical directions of image axes.
+    """
+    directions = []
+    for vec in [row_cos, col_cos, normal]:
+        axis = np.argmax(np.abs(vec))
+        sign = vec[axis]
+        if axis == 0:
+            directions.append('L' if sign >= 0 else 'R')
+        elif axis == 1:
+            directions.append('P' if sign >= 0 else 'A')
+        else:
+            directions.append('S' if sign >= 0 else 'I')
+    return directions
+
+def determine_flips(dcm):
+    iop = np.array(dcm.ImageOrientationPatient)
+    row_cos = iop[:3]
+    col_cos = iop[3:]
+    normal = np.cross(row_cos, col_cos)
+
+    labels = get_axis_labels(row_cos, col_cos, normal)
+    desired = ['P', 'I', 'L']  # x, y, z
+
+    flips = []
+    for label, want in zip(labels, desired):
+        flips.append(label != want)
+    return flips
+
+def load_dicom_metadata(dicom_dir, instance_number):
+    for fname in os.listdir(dicom_dir):
+        fpath = os.path.join(dicom_dir, fname)
+        try:
+            dcm = pydicom.dcmread(fpath, stop_before_pixels=True)
+            if hasattr(dcm, "InstanceNumber") and dcm.InstanceNumber == instance_number:
+                return dcm
+        except:
+            continue
+    raise ValueError(f"Instance {instance_number} not found in {dicom_dir}")
+
+def reorient_pixel_coords(x, y, z, shape, flips):
+    if not flips[0]:  # x: posterior to anterior
+        x = shape[1] - 1 - x
+    if not flips[1]:  # y: inferior to superior
+        y = shape[0] - 1 - y
+    if not flips[2]:  # z: left to right
+        z = shape[2] - 1 - z
+    return x, y, z + 1 
+
+def main(args):
+    df = pd.read_csv(args.csv_input)
+
+    # Only keep rows with relevant conditions
+    df_filtered = df[df["condition"].str.contains("Neural Foraminal Narrowing|Subarticular Stenosis", case=False, na=False)].copy()
+
+    output_rows = []
+
+    for _, row in df_filtered.iterrows():
+        study_id = str(row["study_id"])
+        series_id = str(row["series_id"])
+        instance_number = int(row["instance_number"])
+        x_pix = float(row["x"])
+        y_pix = float(row["y"])
+
+        dicom_dir = os.path.join(args.dicom_root, study_id, series_id)
+
+        # Load and sort all DICOMs in the series
+        dicoms = []
+        for fname in os.listdir(dicom_dir):
+            fpath = os.path.join(dicom_dir, fname)
+            try:
+                dcm = pydicom.dcmread(fpath, stop_before_pixels=True)
+                if hasattr(dcm, "InstanceNumber"):
+                    dicoms.append((dcm.InstanceNumber, dcm))
+            except:
+                continue
+
+        if not dicoms:
+            print(f"Warning: No valid DICOMs found in {dicom_dir}")
+            continue
+
+        dicoms.sort(key=lambda x: x[0])
+        instance_numbers = [inst for inst, _ in dicoms]
+        total_slices = len(instance_numbers)
+
+        try:
+            z_index = instance_numbers.index(instance_number)
+        except ValueError:
+            print(f"Warning: Instance {instance_number} not found in {dicom_dir}")
+            continue
+
+        dcm = dicoms[z_index][1]
+        shape = (int(dcm.Rows), int(dcm.Columns), total_slices)
+        flips = determine_flips(dcm)
+
+        x_new, y_new, z_new = reorient_pixel_coords(x_pix, y_pix, z_index, shape, flips)
+
+        output_rows.append({
+            "study_id": study_id,
+            "series_id": series_id,
+            "instance_number": instance_number,
+            "condition": row["condition"],
+            "level": row["level"],
+            "x_P2A": x_new,
+            "y_I2S": y_new,
+            "z_L2R": z_new
+        })
+
+    # Save as new DataFrame
+    df_out = pd.DataFrame(output_rows)
+    df_out.to_csv(args.csv_output, index=False)
+    print(f"✅ Reoriented CSV saved to: {args.csv_output} (n = {len(df_out)})")
+
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Reorient DICOM pixel coordinates to canonical frame (P→A, I→S, L→R)")
+    parser.add_argument("--dicom_root", type=str, required=True, help="Root directory containing DICOMs organized as study_id/series_id/*.dcm")
+    parser.add_argument("--csv_input", type=str, required=True, help="Input CSV with pixel coordinates and instance numbers")
+    parser.add_argument("--csv_output", type=str, required=True, help="Output CSV with reoriented coordinates")
+
+    args = parser.parse_args()
+    main(args)