radiology-deep-learning

Imaging Deep-Learning Study Design

Use this skill to design (or audit) an imaging deep-learning study so it is reproducible, honestly validated, and CLAIM-compliant. DL imaging papers get torn apart for slice-level splits, patient overlap, test-set tuning, no external validation, and baselines that are too weak to make "deep learning wins" mean anything. This skill encodes the architecture/training choices and the partition hygiene reviewers enforce.

Core stance

• Patient-level everything. Splits, augmentation, and any data-dependent step respect the patient boundary — slices/lesions/sequences/timepoints from one patient never span sets.
• Right capacity for the data. Small cohorts → transfer learning, self-supervised pretraining, strong simple baselines, heavy augmentation, and nested CV — not a giant model trained from scratch on 200 images.
• Beat a real baseline. "DL is better" needs a fair comparator: a radiomics/clinical model, a strong simpler network, or radiologists — tuned as carefully as the proposed model.
• Inputs declared. State exactly how images, masks, clinical variables, text, and molecular data enter the model (channels, crops, fusion point), and how missing modalities are handled.
• External validation is the headline, not a footnote. Internal CV alone is weak; freeze the pipeline and validate on an unseen site/period (→ radiology-design/validation-strategy).
• Report calibration + utility, failure cases, and CIs — not just AUC/Dice (→ radiology-stats).
• Explain, quantify uncertainty, and stress-test. A high-AUC model with no interpretability, no confidence estimate, and no robustness check is under-built for a high-impact venue — RQS 2.0 (2025) scores explainability/fairness directly, and reviewers increasingly ask (→ interpretability-uncertainty.md).
• Integrity. Never invent performance, training curves, or hyperparameters; mark what must be run.

When to use

• "Design a CNN/Transformer/3D/segmentation/detection/prognostic imaging model." / "影像深度学习课题设计。"
• "Transfer learning vs self-supervised vs from scratch for my cohort size?"
• "How should images + clinical + pathology/text enter the model (multimodal fusion)?"
• "Augmentation, class imbalance, hyperparameter search, baselines — how to set up?"
• "Audit my DL Methods for slice-level leakage / patient overlap / test-set tuning."