| name | radiology-deep-learning |
| description | Design and audit imaging deep-learning studies to Radiology (RSNA) / CLAIM 2024 standard, or to Nature-portfolio / FUTURE-AI trustworthy-AI standard — architecture choice (2D/2.5D/3D CNN, Transformer/ViT, segmentation/detection nets, prognostic models), transfer learning vs self-supervised pretraining vs training from scratch, how images/masks/clinical/text/molecular inputs enter the model, data splitting and augmentation, class imbalance, hyperparameter search, baselines, external validation, interpretability/explainability (Grad-CAM, SHAP, attention), uncertainty quantification (MC dropout, ensembles, conformal prediction), and robustness/OOD testing — with patient-level partition hygiene throughout. Use when the user plans or reviews a CNN/Transformer/3D/segmentation/detection/foundation/multimodal imaging model, mentions transfer learning, self-supervised, nnU-Net, ViT, data augmentation, class imbalance, explainability, uncertainty, robustness, or "影像深度学习/深度学习模型". Produces a model+training+validation design, a leakage audit, and Methods text. Never fabricates performance or training details. |
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."
When to open extra files
| File | Open when |
|---|
| references/architecture-choice.md | Choosing 2D/2.5D/3D CNN, Transformer/ViT, segmentation/detection/prognostic heads; foundation models; capacity vs cohort size |
| references/training-protocol.md | Transfer/SSL/from-scratch, splits, augmentation, class imbalance, loss/optimizer/schedule, hyperparameter search, checkpointing, seeds |
| references/multimodal-inputs.md | How images/masks/clinical/text/molecular inputs enter the model; fusion strategies; missing-modality handling |
| references/dl-leakage-audit.md | The DL-specific leakage/validity checklist reviewers weaponise |
| references/interpretability-uncertainty.md | Explainability (Grad-CAM/SHAP/attention) reported without overclaiming; uncertainty quantification (MC dropout, ensembles, conformal prediction); robustness/OOD testing; FUTURE-AI framing |
| references/foundation-models-trustworthy-ai.md | Foundation models, ViT, SSL, VLM/report generation, 3D radiology models, adapter/LoRA, UQ/XAI/causal robustness, or deployment-grade trustworthy AI |
Workflow
- Confirm the design (reuse
radiology-design) — task, endpoint, unit (patient-level),
cohorts, validation type, realistic capacity given n.
- Choose architecture (architecture-choice.md) — dimension, family, task head; justify
capacity vs cohort size; pick the baseline(s) to beat.
- For foundation/trustworthy-AI designs, open
foundation-models-trustworthy-ai.md
and specify pretraining/adaptation, baseline ladder, calibration, UQ, XAI stability,
OOD/robustness, fairness, and shortcut/confounder audits.
- Define inputs (multimodal-inputs.md) — channels/crops/fusion; missing-modality rule;
leakage-safe use of masks and clinical/text/molecular data.
- Set training (training-protocol.md) — transfer/SSL/scratch, patient-level splits,
augmentation, imbalance handling, loss/optimizer/schedule, nested-CV hyperparameter search,
seeds, checkpoint selection (on validation, never test).
- Validate — internal (patient-level CV) + external/temporal/geographic with the pipeline
frozen; report discrimination, calibration, utility, failure cases, CIs (→ radiology-stats).
- Explain + quantify uncertainty + stress-test (interpretability-uncertainty.md) — pick a
method matched to the architecture; report bounded, with failure cases, not only flattering
examples.
- Audit leakage (dl-leakage-audit.md) and write Methods to CLAIM 2024.
Output contract
Model design — architecture, task head, capacity rationale, baseline(s).
Input spec — how each modality enters; fusion point; missing-modality handling.
Training protocol — pretraining strategy, splits, augmentation, imbalance, loss/optim/
schedule, hyperparameter search, seeds, checkpoint rule — reproducibly.
Validation plan — internal + external; metrics incl. calibration/utility + failure cases.
Interpretability & uncertainty — method, parameters, quantitative check, and the bounded
claim it supports (→ interpretability-uncertainty.md).
Trustworthiness modules — for foundation/VLM/deployment-grade models: calibration,
UQ, OOD/fairness, XAI stability, shortcut/confounder audit.
Leakage audit — pass/fail per item + fix.
Methods paragraph — CLAIM-aligned prose (+ 待确认 for Chinese authors).
Quality bar
A good DL design is reproducible from the protocol, splits at the patient level, beats a fair
baseline, validates externally with the pipeline frozen, and reports calibration and failure
cases — not a single AUC from a slice-level split.
Handoffs
- Hand-crafted feature comparison / deep-feature extraction context →
radiology-radiomics.
- Segmentation/detection ground-truth mask SOP and reader reproducibility →
radiology-annotation.
- CLAIM/TRIPOD+AI audit, FUTURE-AI/TRIPOD-LLM edge cases →
radiology-reporting.
- Metrics, CIs, DeLong, calibration, MRMC, sample size →
radiology-stats.
- Validation-type design (external/temporal/multi-center) →
radiology-design.
- Biological interpretation of deep features →
radiology-radiogenomics.
- Reader study / prospective deployment / monitoring for drift →
radiology-translation.
- Figures (architecture, ROC, calibration, Grad-CAM, uncertainty plots) →
radiology-figure.
- Reframing this as a funding proposal instead of / alongside a paper →
radiology-grant.