""" Band Content Statistics - companion to border_ablation.py ========================================================== border_ablation.py found that masking the top/bottom 20% bands drops glioma accuracy by ~30pp while pituitary/notumor are unaffected. Two hypotheses: A) Shortcut: glioma decisions use border-specific artifacts. B) Occlusion: glioma images have bright tissue (lesion/brain) extending into those bands, so masking removes real tumor evidence. This script measures image intensity in the top and bottom 20% bands per class on a split. If glioma bands hold much more bright tissue than the other classes', hypothesis B is favored: masking cut through anatomy, not artifacts. Usage (run from backend/): ..\\.venv\\Scripts\\python.exe src\\band_stats.py \ --config configs\\efficientnet_b0.yaml --split val """ import argparse from pathlib import Path from typing import Dict import numpy as np import pandas as pd from utils import load_config from data.dataset import CLASS_TO_IDX def parse_args() -> argparse.Namespace: parser = argparse.ArgumentParser( description="Brightness statistics of border bands per class" ) parser.add_argument("--config", type=str, required=True) parser.add_argument("--split", type=str, default="val", choices=["train", "val", "test"]) parser.add_argument("--band-frac", type=float, default=0.2, help="Band thickness as fraction of image height") return parser.parse_args() def band_stats(image: np.ndarray, band_frac: float) -> Dict[str, float]: """Fraction of 'tissue' (>20 intensity) pixels in top and bottom bands.""" h = image.shape[0] gray = image.mean(axis=2) if image.ndim == 3 else image top_band = gray[: int(h * band_frac)] bottom_band = gray[int(h * (1.0 - band_frac)):] def tissue(band: np.ndarray) -> float: return float(band.mean() / 255.0) return { "tissue_frac_top": float((top_band > 20).mean()), "tissue_frac_bottom": float((bottom_band > 20).mean()), "mean_top": tissue(top_band), "mean_bottom": tissue(bottom_band), } def main() -> None: args = parse_args() config = load_config(args.config) processed_dir = Path.cwd() / config["data"]["processed_dir"] manifest = pd.read_csv(processed_dir / "manifest.csv") df = manifest[manifest["split"] == args.split].reset_index(drop=True) print(f"Split: {args.split} | Images: {len(df)} | Band frac: {args.band_frac}") rows = [] for filepath, class_name in zip(df["filepath"], df["class"]): image = np.load(processed_dir / filepath) stats = band_stats(image, args.band_frac) rows.append({"class": class_name, **stats}) stats_df = pd.DataFrame(rows) agg = stats_df.groupby("class").mean().reindex( [k for k, _ in sorted(CLASS_TO_IDX.items(), key=lambda kv: kv[1])] ) print() print("Mean fraction of image brightness per band (higher = more tissue)") print(agg.to_string(float_format=lambda v: f"{v:6.3f}")) out_path = Path.cwd() / "reports" / "band_stats.csv" agg.to_csv(out_path, float_format="%.5f") print(f"\nSaved: {out_path}") if __name__ == "__main__": main()