| """Step 5: exploratory analysis, quality flags and shortcut (confound) checks. |
| |
| python -m fsbench.eda --bench data/bench --processed data/processed |
| Writes: eda/eda_report.md, eda/*.png, quality_flags in splits_final.csv |
| """ |
| import argparse |
| from pathlib import Path |
|
|
| import matplotlib |
| matplotlib.use("Agg") |
| import matplotlib.pyplot as plt |
| import numpy as np |
| import pandas as pd |
| from PIL import Image |
| from sklearn.ensemble import RandomForestClassifier |
| from sklearn.metrics import balanced_accuracy_score, roc_auc_score |
| from sklearn.model_selection import GroupKFold, cross_val_predict |
|
|
| CLASSES = ["N", "D", "G", "A", "M", "OPENSET"] |
| CLASS_COLOR = {"N": "#2a78d6", "D": "#eb6834", "G": "#1baf7a", "A": "#eda100", "M": "#e87ba4", |
| "OPENSET": "#a3a29c"} |
| INK, INK2, GRID, SURF = "#0b0b0b", "#52514e", "#e4e3df", "#fcfcfb" |
| SPLITS = ["train", "validation", "calibration", "test", "test_ood", "test_openset"] |
| TRIVIAL = ["orig_width", "orig_height", "aspect", "fov_fraction", "brightness", "contrast", "sharpness", |
| "mean_r", "mean_g", "mean_b"] |
|
|
| plt.rcParams.update({"figure.facecolor": SURF, "axes.facecolor": SURF, "axes.edgecolor": GRID, "axes.labelcolor": INK2, |
| "xtick.color": INK2, "ytick.color": INK2, "text.color": INK, "font.size": 9, |
| "axes.spines.top": False, "axes.spines.right": False, "axes.grid": True, "grid.color": GRID, |
| "grid.linewidth": 0.6, "axes.axisbelow": True}) |
|
|
|
|
| def _md(df, floatfmt="{:.3f}"): |
| df = df.reset_index() |
| fmt = lambda v: floatfmt.format(v) if isinstance(v, float) else str(v) |
| return "\n".join(["| " + " | ".join(map(str, df.columns)) + " |", "|" + "---|" * len(df.columns)] + |
| ["| " + " | ".join(fmt(v) for v in r) + " |" for r in df.itertuples(index=False)]) |
|
|
|
|
| def quality_flags(d): |
| flags = pd.Series("", index=d.index) |
| sharp_lo = d.groupby("source").sharpness.transform(lambda s: s.quantile(0.01)) |
| rules = {"low_fov": d.fov_fraction < 0.15, "very_dark": d.brightness < 25, "overexposed": d.bright_clip > 0.25, |
| "blurry_for_source": d.sharpness < sharp_lo, "tiny": d[["orig_width", "orig_height"]].min(axis=1) < 300} |
| for name, m in rules.items(): |
| flags[m.fillna(False)] += name + ";" |
| return flags.str.rstrip(";") |
|
|
|
|
| def fig_composition(u, out): |
| classes = [c for c in CLASSES if c in set(u.label)] |
| tab = pd.crosstab(u.source, u.label).reindex(columns=classes, fill_value=0) |
| side = u.groupby("source").split.first().map(lambda s: "test" if s.startswith("test") else "dev") |
| tab = tab.loc[sorted(tab.index, key=lambda s: (side[s] != "dev", -tab.loc[s].sum()))] |
| fig, ax = plt.subplots(figsize=(8, 0.32 * len(tab) + 1.2)) |
| left = np.zeros(len(tab)) |
| for c in classes: |
| ax.barh(range(len(tab)), tab[c], left=left, color=CLASS_COLOR[c], label=c, height=0.7, |
| edgecolor=SURF, linewidth=1) |
| left += tab[c].to_numpy() |
| ax.set_yticks(range(len(tab)), [f"{s} ({side[s]})" for s in tab.index]) |
| ax.invert_yaxis() |
| ax.set_xscale("symlog", linthresh=100) |
| ax.set_xlabel("images (symlog)") |
| ax.grid(axis="y", visible=False) |
| ax.legend(ncols=4, loc="lower right", frameon=False, title="class") |
| ax.set_title("Images per source and class", loc="left", fontsize=11) |
| fig.tight_layout() |
| fig.savefig(out / "composition.png", dpi=150) |
| plt.close(fig) |
|
|
|
|
| def fig_acquisition(d, out): |
| srcs = d.groupby("source").size().sort_values(ascending=False).index |
| fig, axes = plt.subplots(1, 3, figsize=(12, 0.3 * len(srcs) + 1.5), sharey=True) |
| for ax, col, title in zip(axes, ["orig_width", "brightness", "sharpness"], |
| ["original width (px)", "mean brightness in field", "Laplacian sharpness (log)"]): |
| data = [d.loc[d.source == s, col].dropna().to_numpy() for s in srcs] |
| if col == "sharpness": |
| data = [np.log10(x + 1) for x in data] |
| ax.boxplot(data, vert=False, widths=0.6, showfliers=False, patch_artist=True, |
| boxprops=dict(facecolor="#2a78d6", alpha=0.35, edgecolor="#2a78d6"), |
| medianprops=dict(color=INK), whiskerprops=dict(color=INK2), capprops=dict(color=INK2)) |
| ax.set_title(title, loc="left", fontsize=10) |
| ax.grid(axis="y", visible=False) |
| axes[0].set_yticks(range(1, len(srcs) + 1), srcs) |
| axes[0].invert_yaxis() |
| fig.suptitle("Acquisition differences between sources", x=0.01, ha="left", fontsize=11) |
| fig.tight_layout() |
| fig.savefig(out / "acquisition.png", dpi=150) |
| plt.close(fig) |
|
|
|
|
| def fig_samples(u, processed, out, per=4): |
| rng = np.random.default_rng(0) |
| srcs = sorted(u.source.unique()) |
| fig, axes = plt.subplots(len(srcs), per, figsize=(per * 1.6, len(srcs) * 1.7)) |
| for i, s in enumerate(srcs): |
| sub = u[u.source == s] |
| pick = sub.sample(frac=1, random_state=0).groupby("label").head(per) |
| pick = pick.sample(min(per, len(pick)), random_state=int(rng.integers(1e6))) |
| for j in range(per): |
| ax = axes[i, j] |
| ax.axis("off") |
| if j < len(pick): |
| r = pick.iloc[j] |
| ax.imshow(Image.open(processed / s / f"{r.image_id}.jpg").resize((160, 160))) |
| ax.set_title(f"{s}·{r.label}" if j == 0 else r.label, fontsize=7, loc="left") |
| fig.tight_layout() |
| fig.savefig(out / "samples.png", dpi=110) |
| plt.close(fig) |
|
|
|
|
| def shortcut_analysis(u): |
| """How much do trivial acquisition statistics reveal source and (within source) label?""" |
| X = u[TRIVIAL].fillna(0).to_numpy() |
| rf = lambda: RandomForestClassifier(200, min_samples_leaf=5, n_jobs=-1, random_state=0) |
| k = min(5, u.group_id.nunique()) |
| pred = cross_val_predict(rf(), X, u.source, groups=u.group_id, cv=GroupKFold(k)) |
| src_bacc = balanced_accuracy_score(u.source, pred) |
| rows = [] |
| for s, sub in u.groupby("source"): |
| for c in CLASSES[:-1]: |
| y = (sub.label == c).astype(int) |
| if y.sum() < 10 or (1 - y).sum() < 10: |
| continue |
| kk = min(5, sub.group_id.nunique()) |
| p = cross_val_predict(rf(), sub[TRIVIAL].fillna(0).to_numpy(), y, groups=sub.group_id, |
| cv=GroupKFold(kk), method="predict_proba")[:, 1] |
| rows.append({"source": s, "class_vs_rest": c, "n": len(sub), "auc_trivial_features": roc_auc_score(y, p)}) |
| return src_bacc, 1 / u.source.nunique(), pd.DataFrame(rows) |
|
|
|
|
| def main(argv=None): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--bench", required=True) |
| ap.add_argument("--processed", required=True) |
| a = ap.parse_args(argv) |
| bench, processed = Path(a.bench), Path(a.processed) |
| out = bench / "eda" |
| out.mkdir(exist_ok=True) |
| s = pd.read_csv(bench / "splits.csv", dtype=str, keep_default_na=False) |
| s["split"] = s.split.replace({"val": "validation", "cal": "calibration"}) |
| man = pd.read_csv(bench / "manifest.csv", dtype=str, keep_default_na=False)[["image_id", "width", "height"]] |
| q = pd.read_csv(bench / "quality.csv") |
| d = s.merge(man, on="image_id", how="left").merge(q, on="image_id", how="left") |
| d["orig_width"] = pd.to_numeric(d.width, errors="coerce") |
| d["orig_height"] = pd.to_numeric(d.height, errors="coerce") |
| d["aspect"] = d.orig_width / d.orig_height |
| d["quality_flags"] = quality_flags(d) |
| d.drop(columns=["width", "height"]).to_csv(bench / "splits_final.csv", index=False) |
|
|
| u = d[d.split.isin(SPLITS)].copy() |
| fig_composition(u, out) |
| fig_acquisition(u, out) |
| fig_samples(u, processed, out) |
| src_bacc, chance, sc = shortcut_analysis(u) |
| sc.to_csv(out / "shortcut_auc.csv", index=False) |
|
|
| flagged = u[u.quality_flags != ""] |
| flag_tab = flagged.quality_flags.str.split(";").explode().value_counts().rename("images") |
| acq = u.groupby("source").agg(images=("image_id", "size"), median_w=("orig_width", "median"), |
| median_h=("orig_height", "median"), brightness=("brightness", "median"), |
| sharpness=("sharpness", "median"), fov=("fov_fraction", "median")) |
| lines = ["# FundusShift — EDA & data-quality report", "", |
| f"Released images: {len(u)} from {u.source.nunique()} sources.", "", |
| "", "", "## Acquisition statistics per source", "", _md(acq, "{:.1f}"), |
| "", "", "", |
| "## Shortcut / confound analysis", "", |
| f"A random forest on 10 trivial image statistics (resolution, aspect, field-of-view, brightness, contrast, " |
| f"sharpness, mean RGB) identifies the **source** with balanced accuracy **{src_bacc:.3f}** " |
| f"(chance {chance:.3f}, group 5-fold CV). Source identity is therefore almost free to learn — " |
| "a model can exploit it whenever label priors differ by source.", "", |
| "Within-source label predictability from the same trivial statistics (AUC; 0.5 = no shortcut). " |
| "AUC > 0.8 means classes in that source were captured under different conditions, so in-source accuracy " |
| "overstates disease recognition:", "", _md(sc.sort_values("auc_trivial_features", ascending=False)), "", |
| "## Quality flags (kept, not removed)", "", |
| _md(flag_tab.rename_axis("flag")) if len(flag_tab) else "No image flagged.", "", |
| "## Sample images (after fundus crop)", "", ""] |
| (out / "eda_report.md").write_text("\n".join(lines) + "\n") |
| print(f"source identifiability bacc={src_bacc:.3f} (chance {chance:.3f})") |
| print(sc.sort_values("auc_trivial_features", ascending=False).to_string(index=False)) |
| print(f"flagged {len(flagged)} images: {flag_tab.to_dict()}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
