{ "annotation_reproduction": { "columns": { "area_max": 0.0, "area_mean": 6.821210263296962e-13, "area_median": 0.0, "area_min": 0.0, "area_std": 9.094947017729282e-13, "area_sum": 2.9103830456733704e-10, "bbox_height_mean": 8.526512829121202e-14, "bbox_height_median": 0.0, "bbox_width_mean": 6.394884621840902e-14, "bbox_width_median": 0.0, "n_annotations": 0.0, "n_points_max": 0.0, "n_points_mean": 0.0, "n_points_median": 0.0, "n_points_min": 0.0, "n_points_sum": 0.0 }, "max_abs_diff": 2.9103830456733704e-10, "max_abs_diff_sql": 0.0, "note": "pandas reference helper and the SQL aggregation used for synthetic recomputation are each compared with lc.v1_image_annotation_stats on a deterministic bounded sample; the focused test checks the pandas helper exhaustively on a fixture", "sampled_images": 500, "sql_columns": { "area_max": 0.0, "area_mean": 0.0, "area_median": 0.0, "area_min": 0.0, "area_std": 0.0, "area_sum": 0.0, "bbox_height_mean": 0.0, "bbox_height_median": 0.0, "bbox_width_mean": 0.0, "bbox_width_median": 0.0, "n_annotations": 0.0, "n_points_max": 0.0, "n_points_mean": 0.0, "n_points_median": 0.0, "n_points_min": 0.0, "n_points_sum": 0.0 }, "total_images": 5387 }, "feature_schema_version": "1", "feature_selection": { "all_missing_handling": "a feature absent for every fit row of a model is dropped for that model; remaining nulls are imputed with the model's fit-only median", "constant_in_fit": [ "width", "height", "image_area_pixels" ], "dropped_all_missing_global": [], "excluded_features": { "annotation_join_status": "join bookkeeping", "cell_line": "acquisition/grouping context", "crop": "acquisition context", "dimensions_match": "audit bookkeeping", "dimensions_status": "audit bookkeeping", "field": "acquisition context", "file_name": "filename (identifier)", "filename_parsed": "metadata flag", "has_annotations": "derived from n_annotations", "has_image_file": "availability flag", "image_features_available": "computation status", "image_features_status": "computation status", "image_id": "original COCO id (identifier)", "is_alternate_version": "version bookkeeping", "is_cross_split_overlap": "version bookkeeping", "local_height": "local TIFF dimension check (nullable)", "local_width": "local TIFF dimension check (nullable)", "metadata_complete": "metadata flag", "original_filename": "filename (identifier)", "record_key": "row key", "record_position": "row key", "records_for_filename": "version bookkeeping", "split": "split identity, not an image property", "splits_for_filename": "version bookkeeping", "time_hours": "acquisition context", "timestamp": "acquisition context", "versions_in_split": "version bookkeeping", "well": "acquisition context", "well_col": "acquisition context", "well_row": "acquisition context", "zero_annotation": "derived from n_annotations" }, "global_columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "model_features": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "robust_features": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_max" ], "zero_mad_handling": "robust baseline: a zero MAD falls back to the fit-only global MAD; if that is also zero the feature is skipped for that row" }, "fitted_cell_lines": [ "A172", "BT474", "BV2", "Huh7", "MCF7", "SHSY5Y", "SKOV3", "SkBr3" ], "global_fallback_rule": "a cell line with fewer than 20 fit images uses the documented global model fitted on all fit rows; no model is ever fitted on validation, test, held-out or known-case rows", "inputs": { "feature_manifest_sha256": "bb3c676cefca78cf8e145b1e68d5074f61745355989496109aec15034baa4109", "feature_records": 5387, "feature_schema_version": "1", "source_sha256": { "test": "9eb5805ca031cebcd70bab4a4aa0c89d98646c206d333803a8d5b10ea7fe9960", "train": "6788902479bc0ffd2838fb7160a778078f464ea6034ed76bb5a0e40263cb9e78", "val": "eb72a68f5065d9ade1805b496eecdcd14ca275e5f1bc6aae7e07279bddadfdc7" }, "warehouse_manifest_sha256": "d53e986df4389e55adec7b42c51baf0daf0a567cfcaf30996a375d411d2fa85f", "warehouse_schema_version": "1", "warehouse_view_version": 1 }, "limitations": [ "Scores are unusualness indicators; a sparse image may be annotated correctly.", "The nine known cases are a regression exercise, not recall over all bad labels.", "Acquisition groups are a leakage proxy, not proof of biological independence.", "Official test imagery is ranked only for exploration, never for tuning.", "No calibration against real annotation error is claimed or performed.", "The Isolation Forest mixes label statistics with image texture, so it flags unusual images rather than labels that disagree with their image; time and well context are excluded from its features.", "3 of 24 model features are constant across the fit rows (width, height, image_area_pixels) and carry no information." ], "models": { "A172": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 257, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "BT474": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 281, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "BV2": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 256, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "Huh7": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 259, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "MCF7": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 317, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "SHSY5Y": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 302, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "SKOV3": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 197, "kind": "cell_line", "max_samples": { "101": 197, "42": 197, "7": 197 } }, "SkBr3": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 303, "kind": "cell_line", "max_samples": { "101": 256, "42": 256, "7": 256 } }, "global": { "columns": [ "n_annotations", "area_sum", "area_mean", "area_median", "area_min", "area_max", "area_std", "bbox_width_mean", "bbox_width_median", "bbox_height_mean", "bbox_height_median", "n_points_mean", "n_points_median", "n_points_min", "n_points_max", "n_points_sum", "pixel_mean", "pixel_std", "gradient_mean", "pixel_min", "pixel_max", "width", "height", "image_area_pixels" ], "dropped_all_missing": [], "fit_rows": 2172, "kind": "global_fallback", "max_samples": { "101": 256, "42": 256, "7": 256 } } }, "partition": { "caveat": "practical leakage proxy only, not proof of biological independence; with 4-6 training acquisition groups per cell line a held-out group is a coarse holdout", "cell_lines": { "A172": { "cell_line": "A172", "fit_groups": [ "A172|A7|field1", "A172|A7|field2", "A172|D7|field1", "A172|D7|field2" ], "holdout_groups": [ "A172|B7|field1", "A172|B7|field2" ], "n_fit_images": 257, "n_groups": 6, "n_holdout_images": 129, "n_known_case_records_excluded": 2, "partition_key": "A172" }, "BT474": { "cell_line": "BT474", "fit_groups": [ "BT474|A3|field1", "BT474|A3|field2", "BT474|C3|field1", "BT474|C3|field2" ], "holdout_groups": [ "BT474|B3|field1", "BT474|B3|field2" ], "n_fit_images": 281, "n_groups": 6, "n_holdout_images": 147, "n_known_case_records_excluded": 1, "partition_key": "BT474" }, "BV2": { "cell_line": "BV2", "fit_groups": [ "BV2|B4|field1", "BV2|B4|field2", "BV2|D4|field1", "BV2|D4|field2" ], "holdout_groups": [ "BV2|C4|field1", "BV2|C4|field2" ], "n_fit_images": 256, "n_groups": 6, "n_holdout_images": 128, "n_known_case_records_excluded": 2, "partition_key": "BV2" }, "Huh7": { "cell_line": "Huh7", "fit_groups": [ "Huh7|A10|field1", "Huh7|A10|field2", "Huh7|A11|field1" ], "holdout_groups": [ "Huh7|A11|field2" ], "n_fit_images": 259, "n_groups": 4, "n_holdout_images": 93, "n_known_case_records_excluded": 0, "partition_key": "Huh7" }, "MCF7": { "cell_line": "MCF7", "fit_groups": [ "MCF7|E4|field1", "MCF7|E4|field2", "MCF7|G4|field1", "MCF7|G4|field2" ], "holdout_groups": [ "MCF7|F4|field1", "MCF7|F4|field2" ], "n_fit_images": 317, "n_groups": 6, "n_holdout_images": 152, "n_known_case_records_excluded": 0, "partition_key": "MCF7" }, "SHSY5Y": { "cell_line": "SHSY5Y", "fit_groups": [ "SHSY5Y|B10|field1", "SHSY5Y|B10|field2", "SHSY5Y|D10|field1", "SHSY5Y|D10|field2" ], "holdout_groups": [ "SHSY5Y|C10|field1", "SHSY5Y|C10|field2" ], "n_fit_images": 302, "n_groups": 6, "n_holdout_images": 146, "n_known_case_records_excluded": 1, "partition_key": "SHSY5Y" }, "SKOV3": { "cell_line": "SKOV3", "fit_groups": [ "SKOV3|G4|field1", "SKOV3|G4|field2", "SKOV3|H4|field1" ], "holdout_groups": [ "SKOV3|H4|field2" ], "n_fit_images": 197, "n_groups": 4, "n_holdout_images": 69, "n_known_case_records_excluded": 0, "partition_key": "SKOV3" }, "SkBr3": { "cell_line": "SkBr3", "fit_groups": [ "SkBr3|E3|field1", "SkBr3|E3|field2", "SkBr3|H3|field1", "SkBr3|H3|field2" ], "holdout_groups": [ "SkBr3|F3|field1", "SkBr3|F3|field2" ], "n_fit_images": 303, "n_groups": 6, "n_holdout_images": 146, "n_known_case_records_excluded": 0, "partition_key": "SkBr3" } }, "counts": { "canonical_filenames": 5239, "canonical_filenames_no_eligible_version": 0, "distinct_filenames_total": 5239, "excluded_known_case": 6, "excluded_unusable": 0, "excluded_unusable_filenames": 0, "fit": 2172, "held_out": 1010, "test": 1512, "validation": 539 }, "policy": "group related images by (cell_line, well, field); keep time points, crops and alternate annotation versions together; split each cell line's eligible training groups deterministically ~75/25 with at least one fit and one held-out group when possible" }, "protocol": { "budgets": [ 10, 25, 50 ], "default_seed": 42, "isolation_forest": { "contamination": "auto", "max_samples": "min(256, n_fit)", "n_estimators": 200, "n_jobs": 1, "score_transform": "priority = -score_samples (higher = more unusual)" }, "min_cell_line_fit_images": 20, "partition_holdout_fraction": 0.25, "partition_seed": 42, "preprocessing": "sklearn Pipeline(SimpleImputer(median), RobustScaler, IsolationForest), fitted on fit rows only", "sensitivity_seeds": [ 7, 42, 101 ] }, "question": "At the same review budget, does an Isolation Forest ranking offer useful prioritisation compared with simple baselines? A negative result is valid.", "ranking_schema_version": "1", "score_semantics": "Scores indicate unusualness, never the probability that a label is wrong. Priority is a fit-only empirical percentile in [0, 1] and is a relative review priority, not an error probability.", "stage": "ranking" }