RSNA Continual Learning

Continual-learning experiments on the RSNA 2023 Abdominal Trauma Detection dataset. We compare Baseline (fine-tuning), EWC, Experience Replay, and EWC + Replay on a small CNN, and report average accuracy and forgetting.

Tags: continual-learning · catastrophic-forgetting · ewc · experience-replay · learning-without-forgetting · multiple-instance-learning · medical-imaging · ct · rsna · pytorch

Repository layout

.
├── src/                       # experiment code
│   ├── exp_class_incremental.py   # Exp 1: 2-task class-incremental (headline)
│   ├── exp_window_3task_v2.py     # Exp 2: 3-task window domain-incremental
│   ├── exp_window_3task_v3.py     # later window variant (near-chance, kept for ref)
│   ├── exp_improved.py            # mods 1-6 (ResNet-18, λ-sweep, balanced replay,
│   │                              #   herding, LwF, patient agg.) — 2-task, diagnostic
│   ├── exp_improved_v2.py         # 3-window full fine-tune (slice-level; near-chance)
│   ├── exp_mil.py                 # patient-level attention-MIL (best CL result)
│   ├── tune_mil.py                # validation-AUC hyperparameter sweep for exp_mil
│   ├── config.py                  # central experiment configuration + presets
│   ├── utils.py                   # helpers
│   └── quickstart.py              # scaffold (does not load data)
├── notebooks/                 # interactive walkthroughs (local only; empty on GitHub)
├── report/                    # report
│   ├── figures/                   # committed vector figures (PDF) — the only tracked report files
│   └── legacy/                    # earlier plain-text reports (local only; empty on GitHub)
├── logs/                      # run logs / CSVs / matrices (local only; empty on GitHub)
└── data/                      # dataset (local only; empty on GitHub)

Data layout

Expected on disk under data/:

data/
  RSNA2023ProcessedImages/<patient_id>/<series_id>/<instance>.png
  train.csv               # labels; uses the `any_injury` column
  image_level_labels.csv

Running the experiments

conda activate medical_ml

# Smoke test (limits patients/images/iterations)
$env:DEBUG_RUN = "1"; python src/exp_class_incremental.py

# Full run
Remove-Item Env:DEBUG_RUN -ErrorAction SilentlyContinue
python src/exp_class_incremental.py

Window-based variants: python src/exp_window_3task_v2.py.

Improved experiment (next iteration)

src/exp_improved.py implements the six modifications derived from the report analysis: (1) pretrained ResNet-18 backbone, (2) EWC λ sweep {10, 50, 100, 500, 1000}, (3) balanced replay loss, (4) larger buffer + herding exemplar selection, (5) LwF knowledge distillation, (6) patient/series-level label aggregation. It is self-contained (no full-run side effects on import).

$env:DEBUG_RUN = "1"; python src/exp_improved.py     # fast smoke test (verified)
Remove-Item Env:DEBUG_RUN; python src/exp_improved.py # full run (multi-hour, GPU)