Radiology Guardians: Adaptive AI for SIIM COVID-19 Detection

Radiology Guardians is an end-to-end research workspace for learning, validating, and packaging thoracic imaging detectors tailored to the SIIM-FISABIO-RSNA challenge. The codebase orchestrates classification, segmentation, and localization models while tracking every artifact needed for reproducible experiments.

Highlights

• Full training recipes covering multi-task classification, lung localization, and opacity detection.
• Seamless pseudo-labeling loops that recycle confident predictions and strengthen supervision.
• Modular data preparation utilities that transform raw DICOM studies into curated model inputs.
• Built-in visual diagnostics, including a demo notebook and exported flowcharts, to accelerate review cycles.

Environment Setup

The project was originally developed on multi-GPU Linux hosts with CUDA-enabled PyTorch. For local workstations:

• Recommended: Python 3.8+, CUDA 11.x compatible GPUs, 64 GB RAM minimum when generating pseudo-labels.
• Create an isolated environment and install dependencies:

  conda create -n radiology-guardians python=3.8
  conda activate radiology-guardians
  pip install -r requirements.txt
  

• Optional native libraries such as GDCM may be required to read certain DICOM encodings. Install them through your OS package manager before running conversion scripts.

Data Preparation

1. Obtain the SIIM-FISABIO-RSNA COVID-19 Detection competition data and place the extracted structure under dataset/siim-covid19-detection.
2. (Optional) Augment with external studies such as RSNA Pneumonia, VinBigData, NIH ChestXray14, CheXpert, PadChest, or pneumothorax datasets. Store each set within dataset/external_dataset/<dataset_name>.
3. Use the preprocessing utilities in src/prepare to convert DICOM files to PNG/JPEG, align annotations, and generate stratified folds:

   cd src/prepare
   python dicom2image_siim.py
   python kfold_split.py
   python prepare_siim_annotation.py
   

4. Execute analogous scripts for any external source you enable (for example dicom2image_pneumonia.py, prepare_vinbigdata.py, refine_data.py). When integrating additional public test predictions for pseudo-labeling, copy the derived lung crops into dataset/lung_crop.
5. Confirm your folder tree matches the reference outline in dataset/dataset_structure.txt.

Pipeline Overview

The processing graph illustrates how raw studies feed into dedicated model families (classification, lung detection, opacity detection) before merging into ensemble inference and pseudo-labeling cycles.

Training Playbooks

Multi-Head Classification with Segmentation Heads

1. Navigate to src/classification_aux.
2. Pretrain backbones on the chosen external sets using train_chexpert_chest14.sh and train_rsnapneu.sh.
3. Fine-tune on the SIIM challenge data via train_siim.sh.
4. Produce refined soft labels and masks using generate_pseudo_label.sh <checkpoint_dir>.
5. Iterate pseudo-labeling with train_pseudo.sh <old_ckpt> <new_ckpt> until validation metrics plateau.
6. Evaluate single models or ensembles with evaluate.py --cfg <config_file> --num_tta <count>.

Classification head performance snapshot (mAP@0.5) across negative, typical, indeterminate, atypical findings:

Model Variant	No TTA	8× TTA
SeResNet152d + UNet decoder	0.575	0.584
EfficientNet-B5 + DeeplabV3+	0.583	0.592
EfficientNet-B6 + LinkNet	0.580	0.587
EfficientNet-B7 + UNet++	0.589	0.595
Geometric ensemble	0.595	0.598

Lung Field Localization

1. Switch to src/detection_lung_yolov5.
2. Download COCO initialization weights by running cd weights && bash download_coco_weights.sh.
3. Train folds through bash train.sh.
4. Average precision summary:

   Metric	Fold 0	Fold 1	Fold 2	Fold 3	Fold 4	Mean
   mAP@0.5:0.95	0.921	0.931	0.926	0.923	0.922	0.925
   mAP@0.5	0.997	0.998	0.997	0.996	0.998	0.997

Opacity Detection

YOLOv5x6 @ 768 px

cd src/detection_yolov5
cd weights && bash download_coco_weights.sh && cd ..
bash train_rsnapneu.sh
bash train_siim.sh
bash generate_pseudo_label.sh
bash warmup_ext_dataset.sh
bash train_final.sh

EfficientDet-D7 @ 768 px

cd src/detection_efffdet
bash train_rsnapneu.sh
bash train_siim.sh
bash generate_pseudo_label.sh
bash warmup_ext_dataset.sh
bash train_final.sh

Faster R-CNN FPN (ResNet101d & ResNet200d)

cd src/detection_fasterrcnn
CUDA_VISIBLE_DEVICES=0,1,2,3 python train_chexpert_chest14.py --steps 0 1 --cfg configs/resnet200d.yaml
CUDA_VISIBLE_DEVICES=0,1,2,3 python train_chexpert_chest14.py --steps 0 1 --cfg configs/resnet101d.yaml
CUDA_VISIBLE_DEVICES=0 python train_rsnapneu.py --cfg configs/resnet200d.yaml
CUDA_VISIBLE_DEVICES=0 python train_rsnapneu.py --cfg configs/resnet101d.yaml
CUDA_VISIBLE_DEVICES=0 python train_siim.py --cfg configs/resnet200d.yaml --folds 0 1 2 3 4 --SEED 123
CUDA_VISIBLE_DEVICES=0 python train_siim.py --cfg configs/resnet101d.yaml --folds 0 1 2 3 4 --SEED 123
CUDA_VISIBLE_DEVICES=0 python warmup_ext_dataset.py --cfg configs/resnet200d.yaml
CUDA_VISIBLE_DEVICES=0 python warmup_ext_dataset.py --cfg configs/resnet101d.yaml
CUDA_VISIBLE_DEVICES=0 python train_final.py --cfg configs/resnet200d.yaml
CUDA_VISIBLE_DEVICES=0 python train_final.py --cfg configs/resnet101d.yaml

If you encounter file descriptor limits on some platforms, enable the optional snippet inside the training scripts to raise the open-file threshold.

Pseudo-Label Filtration

Use src/detection_make_pseudo to aggregate predictions across detectors:

cd src/detection_make_pseudo
python make_pseudo.py
python make_annotation.py

Default thresholds retain studies with negative probability below 0.3 and at least one positive phenotype above 0.7, keeping the two highest-confidence boxes per image. Expect memory usage that peaks around 128 GB when combining every fold and detector.

Detection Benchmarks

Detector	mAP@0.5 (TTA)
YOLOv5x6 768	0.580
EfficientDet-D7 768	0.594
Faster R-CNN ResNet200d 768	0.592
Faster R-CNN ResNet101d 1024	0.596

Evaluation and Reporting

• src/demo_notebook/demo.ipynb visualizes predictions, masks, and box overlays for sanity checks.
• src/classification_aux/evaluate.py consolidates per-fold metrics and renders CSV summaries.
• src/detection_yolov5/test.py, src/detection_efffdet/validate.py, and src/detection_fasterrcnn/utils.py provide detector-specific evaluation hooks with optional test-time augmentation.
• Final challenge submissions combined ensemble logits and weighted boxes, yielding 0.658 public leaderboard and 0.635 private leaderboard scores.

Repository Map

• dataset/ – raw and processed imagery plus fold metadata.
• src/prepare/ – ingestion, conversion, and deduplication tools.
• src/classification_aux/ – hybrid classification/segmentation models, configs, and pseudo-label scripts.
• src/detection_* – detectors (YOLOv5, EfficientDet, Faster R-CNN) with training, inference, and warmup utilities.
• src/detection_make_pseudo/ – pseudo-label curation utilities.
• images/ – diagrams and marketing assets.

License

This project is released under the terms stated in LICENSE. Please review the file before redistributing models or derived datasets.