No rows.
" + parts = ["| {html.escape(field)} | ") + parts.append("
|---|
| {html.escape(str(value))} | ") + parts.append("
Tile records: {len(tile_rows)}. Slide records: {len(slide_rows)}. " + f"Skipped heatmap slide folders: {len(skipped)}.
", + "Subject: {html.escape(report_summary['email_subject'])}
", + "" + html.escape("\n".join(report_summary["email_body"])) + "",
+ "" + f"GT blast %: {html.escape(str(row['blast_percent_gt']))}; " + f"top proxy: {format_number(row['top_enrichment_proxy_mean'])}; " + f"delta: {format_number(row['top_minus_bottom_enrichment_proxy'])}; " + f"technical concern: {row['technical_concern_score']}" + "
" + ) + parts.append( + f"" + f"Full table is in tile_metrics.csv; "
+ "showing first 100 rows here.
No rows.
" + parts = ["| {html.escape(field)} | ") + parts.append("
|---|
| {html.escape(str(value))} | ") + parts.append("
" + f"Candidate mean: {format_number(row['all_candidate_count_mean'])}; " + f"focus mean: {format_number(row['zoom_laplacian_var_mean'])}; " + f"background mean: {format_number(row['zoom_background_fraction_mean'])}; " + f"concern score: {row['technical_concern_score']}" + "
" + ) + parts.append( + f"" + f"Full table is in tile_metrics.csv; "
+ "showing first 40 rows here.
SVG not found
" + return path.read_text() + + +def load_clini_and_slide(): + """Load clinical and slide tables.""" + clini_df = pd.read_csv(CLINI_TABLE) + slide_df = pd.read_csv(SLIDE_TABLE) + fname_to_sample = dict(zip(slide_df["FILENAME"], slide_df["SAMPLE_ID"])) + return clini_df, slide_df, fname_to_sample + + +# ============================================================================= +# Regression experiment helpers +# ============================================================================= + + +def load_regression_stats(exp): + """Load regression statistics.""" + stats_dir = exp["base_dir"] / "statistics" + name = exp["name"] + agg_path = stats_dir / f"{name}_regression-stats_aggregated.csv" + agg = pd.read_csv(agg_path, index_col=0) + return { + "r2": float(agg.loc["r2_score", "mean"]), + "pearson_r": float(agg.loc["pearson_r", "mean"]), + "pearson_p": float(agg.loc["pearson_p", "mean"]), + "mae": float(agg.loc["mae", "mean"]), + "rmse": float(agg.loc["rmse", "mean"]), + "count": int(float(agg.loc["count", "mean"])), + } + + +def make_regression_section(exp): + """Generate the HTML section for a regression experiment.""" + stats = load_regression_stats(exp) + stats_dir = exp["base_dir"] / "statistics" + scatter_svg_path = stats_dir / "plots" / "fold_patient-preds_scatter.svg" + + html = f""" +{exp["description"]}
+ +Each point represents one patient (test-set prediction from the respective crossval fold). +The dashed line is the identity line (perfect prediction). Pearson r = {stats["pearson_r"]:.3f} +indicates moderate-strong correlation. Note the R² = {stats["r2"]:.3f} is near zero due to the +heavily skewed distribution (many samples near 0%, few at high %).
+""" + return html + + +# ============================================================================= +# Classification experiment helpers +# ============================================================================= + + +def load_classification_stats(exp): + """Load classification statistics (works for both binary and multi-class).""" + stats_dir = exp["base_dir"] / "statistics" + name = exp["name"] + agg = pd.read_csv(stats_dir / f"{name}_categorical-stats_aggregated.csv") + ind = pd.read_csv(stats_dir / f"{name}_categorical-stats_individual.csv") + return agg, ind + + +def make_classification_section(exp): + """Generate the HTML section for a classification experiment.""" + agg, ind = load_classification_stats(exp) + stats_dir = exp["base_dir"] / "statistics" + true_class = exp["true_class"] + name = exp["name"] + categories = exp["categories"] + + # Parse aggregated stats + # agg columns: unnamed_class, roc_auc_score (mean), roc_auc_score (95%_low), roc_auc_score (95%_high), + # average_precision_score (mean, low, high), f1_score (mean, low, high), count + # Skip header row (row 0 is the sub-header) + class_stats = {} + for _, row in agg.iterrows(): + cls = str(row.iloc[0]) + if cls in ("", "nan") or cls.startswith("Unnamed"): + continue + class_stats[cls] = { + "auroc_mean": float(row.iloc[1]), + "auroc_low": float(row.iloc[2]), + "auroc_high": float(row.iloc[3]), + "auprc_mean": float(row.iloc[4]), + "auprc_low": float(row.iloc[5]), + "auprc_high": float(row.iloc[6]), + "f1_mean": float(row.iloc[7]), + "f1_low": float(row.iloc[8]), + "f1_high": float(row.iloc[9]), + "count": int(float(row.iloc[10])), + } + + html = f""" +{exp["description"]}
+ +| Fold | Class | Count | AUROC | AUPRC | F1 | p-value | " + html += "
|---|---|---|---|---|---|---|
| {fold} | {cls} | " + ) + html += ( + f"{count} | {auroc:.4f} | {ap:.4f} | {f1:.4f} | " + ) + html += f"{pval:.2e} |
{slide["stem"]}
+{slide["stem"]}
++ Generated {now} — STAMP Framework, UNI2 features, ViT MIL model +
+ ++Four prediction targets from AML bone marrow aspirates: treatment response (RESPONSE_CR) and three +targets derived from the BLAST_PERCENT column. +All experiments use the same dataset (489 samples, 69 biological patients), feature extractor (UNI2, 3072-dim), +and MIL architecture (ViT, 2-layer, 8-head, 512-dim, max_epochs=32, patience=16). +
+ +| Experiment | +Task | +Classes | +Primary Metric | +Result | +
|---|---|---|---|---|
| RESPONSE_CR | +Binary Classification | +yes (CR) / no | +AUROC | +{rc_auroc:.3f} (AUPRC={rc_auprc:.3f}) | +
| BLAST_PERCENT | +Regression | +Continuous (0–91%) | +Pearson r | +{bp_stats["pearson_r"]:.3f} (MAE={bp_stats["mae"]:.1f}%) | +
| BLAST_SEVERITY | +3-class Classification | +high / intermediate / low | +AUROC (per-class) | +high={bs_classes.get("high", 0):.3f}, int={bs_classes.get("intermediate", 0):.3f}, low={bs_classes.get("low", 0):.3f} | +
| HIGH_BLAST | +Binary Classification | +yes (≥20%) / no | +AUROC | +{hb_auroc:.3f} (AUPRC={hb_auprc:.3f}) | +
+Key finding: The binary HIGH_BLAST classifier achieves the best discriminative performance +(AUROC=0.929), suggesting the model can reliably identify high-blast cases. RESPONSE_CR also shows +strong performance (AUROC=0.870), confirming the model captures morphological features predictive of +treatment response. The 3-class BLAST_SEVERITY model shows strong performance for high (0.912) and +low (0.864) classes but struggles with the intermediate class (0.783), consistent with the small +sample size (n=55) and the inherent difficulty of distinguishing 5–19% blast ranges. The regression +model shows moderate correlation (r=0.680) but high RMSE (24.5%) driven by the skewed distribution. +
+SVG not found
" + return path.read_text() + + +def load_stats(): + """Load the per-fold and aggregated statistics CSVs.""" + individual = pd.read_csv(STATS_DIR / "RESPONSE_CR_categorical-stats_individual.csv") + aggregated = pd.read_csv(STATS_DIR / "RESPONSE_CR_categorical-stats_aggregated.csv") + return individual, aggregated + + +def collect_heatmap_slides(): + """Collect all heatmap slide info with predictions and ground truth.""" + slide_df = pd.read_csv(SLIDE_TABLE) + fname_to_sample = dict(zip(slide_df["FILENAME"], slide_df["SAMPLE_ID"])) + + # Load all predictions + all_preds = {} + for split_i in range(5): + pred_file = CROSSVAL_DIR / f"split-{split_i}" / "patient-preds.csv" + if not pred_file.exists(): + continue + preds = pd.read_csv(pred_file) + for _, row in preds.iterrows(): + all_preds[row["SAMPLE_ID"]] = { + "split": split_i, + "pred": row["pred"], + "prob_yes": row["RESPONSE_CR_yes"], + "prob_no": row["RESPONSE_CR_no"], + "gt": row["RESPONSE_CR"], + } + + slides = [] + for split_dir in sorted(HEATMAP_DIR.glob("split-*")): + split_i = int(split_dir.name.split("-")[1]) + for slide_dir in sorted(split_dir.iterdir()): + if not slide_dir.is_dir(): + continue + stem = slide_dir.name + h5_name = stem + ".h5" + sample_id = fname_to_sample.get(h5_name, "unknown") + info = all_preds.get(sample_id, {}) + + overview_path = slide_dir / "plots" / f"overview-{stem}.png" + if not overview_path.exists(): + continue # Skip slides with no overview (e.g., 1-tile failures) + + # Collect top/bottom tiles + tiles_dir = slide_dir / "tiles" + top_tiles = ( + sorted( + [ + t + for ext in ("*.png", "*.jpg", "*.jpeg") + for t in tiles_dir.glob(f"top_*{ext}") + ] + )[:4] + if tiles_dir.exists() + else [] + ) + bottom_tiles = ( + sorted( + [ + t + for ext in ("*.png", "*.jpg", "*.jpeg") + for t in tiles_dir.glob(f"bottom_*{ext}") + ] + )[:4] + if tiles_dir.exists() + else [] + ) + + slides.append( + { + "stem": stem, + "sample_id": sample_id, + "split": split_i, + "gt": info.get("gt", "?"), + "pred": info.get("pred", "?"), + "prob_yes": info.get("prob_yes", 0), + "prob_no": info.get("prob_no", 0), + "overview_path": overview_path, + "top_tiles": top_tiles, + "bottom_tiles": bottom_tiles, + "correct": info.get("gt") == info.get("pred"), + } + ) + + return slides + + +def generate_html(): + individual, aggregated = load_stats() + slides = collect_heatmap_slides() + + # Group slides by ground truth + yes_slides = [s for s in slides if s["gt"] == "yes"] + no_slides = [s for s in slides if s["gt"] == "no"] + + # Extract mean AUROC from aggregated CSV + # The columns are: class, mean, 95%_low, 95%_high (for roc_auc_score) + mean_auroc = float(aggregated.iloc[1, 1]) # yes class, roc_auc_score mean + auroc_low = float(aggregated.iloc[1, 2]) + auroc_high = float(aggregated.iloc[1, 3]) + mean_auprc = float(aggregated.iloc[1, 4]) + + html = f""" + + + + +Generated {datetime.now().strftime("%Y-%m-%d %H:%M")} — STAMP Framework
+ ++Task: Binary classification predicting complete remission (CR) response in AML patients from +H&E-stained bone marrow aspirate whole-slide images. Each sample represents a bone marrow biopsy section +(parts A, B, C) from a timepoint (year 19 or 26). The positive class ("yes") indicates blast percentage +below 5% (complete remission), while "no" indicates ≥5% blasts. +
+| Fold | Class | Count | AUROC | AUPRC | F1 | p-value | " + html += "
|---|---|---|---|---|---|---|
| {fold} | {cls} | ' + html += ( + f"{count} | {auroc:.4f} | {ap:.4f} | {f1:.4f} | " + ) + html += f"{pval:.2e} |
{slide["stem"]}
+Review top/bottom attended native tiles and " + "centered zoom crops. Fill answers in review_answers_template.csv.
" + ), + ] + + for slide in slides: + meta = slide["metadata"] + parts.append("| {html.escape(key)} | " + f"{html.escape(str(meta.get(key, '')))} | " + "
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ The STAMP preprocessing pipeline applies two background-rejection filters at tile-extraction time
+(see src/stamp/preprocessing/tiling.py):
brightness_cutoff = 240. Computed at the
+ supertile level on the slide thumbnail. Any region with average grayscale
+ intensity ≥ 240 is treated as glass/empty background and skipped before any
+ tiling occurs.canny_cutoff = 0.02. After tiling, each
+ 224×224 tile is run through OpenCV Canny (thresholds 40 / 100).
+ Tiles with fewer than 2 % edge pixels are deemed texture-less and dropped.For the four slides above, almost every rejected tile fails the +Canny edge test, not the brightness test. Looking at the rejected-tile +gallery makes the cause obvious: the rejected tiles are out of focus +(blurred, soft, or featureless purple/pink). When the smear is blurry, the +Canny detector finds few edges → the tile is classified as low-information and +discarded.
+ +This correlates with the focus-acquisition setting in the file name:
+| Slide | Focus pts | Retention |
|---|---|---|
| 116 26 B (10 Fokuspunkte) | 10 | 76.9 % |
| 127 26 C (5 Fokuspunkte) | 5 | 4.7 % |
| 333 19 A (5 Fokuspunkte) | 5 | 50.0 % |
| 931 19 A (5 Fokuspunkte) | 5 | 39.7 % |
Summary: The drop in retention is not because the slide is empty — +the tissue is there, but the sub-sampled focus stack (5 Fokuspunkte) failed to +keep most of the smear in sharp focus. The Canny filter is doing its job: +removing tiles that would have been useless features for the model. The +real fix is upstream — re-scan the affected slides with the full +10-Fokuspunkte stack.
+For each slide we show the four task-specific heatmaps generated using the crossval split +in which that slide was held out (so the model never saw it during training). +Each panel pair is: left — overview with the top-K relevant tiles outlined; +right — class map (per-tile class assignment) for classification, or a continuous +relevance heatmap for the regression head.
+ +| BLAST_PERCENT | 2.0% |
| RESPONSE_CR | yes |
| BLAST_SEVERITY | low |
| HIGH_BLAST | no |
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ | BLAST_PERCENT | 4.0% |
| RESPONSE_CR | yes |
| BLAST_SEVERITY | low |
| HIGH_BLAST | no |
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ | BLAST_PERCENT | 0.0% |
| RESPONSE_CR | yes |
| BLAST_SEVERITY | low |
| HIGH_BLAST | no |
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ | BLAST_PERCENT | 0.0% |
| RESPONSE_CR | yes |
| BLAST_SEVERITY | low |
| HIGH_BLAST | no |
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ | BLAST_PERCENT | 0.0% |
| RESPONSE_CR | yes |
| BLAST_SEVERITY | low |
| HIGH_BLAST | no |
+ 
+ 