Medical Imaging - Method Selection & Rigor

High stakes + small, leaky-prone datasets. Rigor matters more than architecture.

Decision table

Non-negotiable rigor (this is where most medical ML fails)

• PATIENT-LEVEL split. All images/views/slices of one patient in ONE split. Use the dataset's official split when it exists (often encoded in the ID). Add asserts (data-rigor-and-leakage). Random per-image splitting is the classic fake-accuracy trap.
• Whole-image vs ROI-crop honesty. Cropped-lesion ROI classification is an easier, different problem; ~98-99% papers are usually ROI-crop and/or leaky. Report which one you are doing and never conflate.
• Medical preprocessing: CLAHE for contrast; crop to the organ/breast (largest connected component) to drop background/labels/tape; remove confounders (e.g., pectoral muscle on mammogram MLO - verify the model isn't attending to it via Grad-CAM); intensity normalization (windowing for CT, z-score for MRI). Apply identically at inference (parity).
• Clinical operating point: report sensitivity AND specificity; prefer a high-sensitivity threshold (recall ≥ ~0.90) chosen on validation - missing cancer >> a false alarm. Don't ship a 0.5 or Youden threshold blindly.
• Calibration: temperature-scale probabilities so "78% malignant" is trustworthy for decision support.
• Framing: decision-support / second-reader, NOT autonomous diagnosis; prototype needs clinical + regulatory validation (deployment-explainability).

Realistic ceilings (whole-image, rigorous)

State honest targets from literature-review rather than chasing leaderboard peaks (e.g., CBIS-DDSM whole-image benign-vs-malignant ≈ 0.83 AUC single-view, ~0.86 two-view; higher numbers are typically ROI/leaky). Use accuracy-improvement-loop to push toward the ceiling.