Medical Image Representation Learning

MIRL is an exploratory project on efficient representation learning on medical imagery built by the SMILE lab at UF

Installation

Use the package manager conda to install MIRL.

conda env create -f environment.yml

Model Architectures & Methodology

• Semi-Supervised Learning
  • Unsupervised training stage
    • Momentum Contrast (MoCo)
    • Simple Contrastive Learning of Representations (SimCLR)
    • Momentum Contrast Version 2 (MoCoV2)
  • Finetuning stage (downstream task)
    • Linear Classifier which freezes all backbone encoder weights (enabled with --frozen)
    • Transfer Learning which retrains all layers (default)
• Supervised Learning
  • ResNet-50

Datasets

Name	Image Size	Unlabeled Count	Labeled Count	#Classes	Multilabel
RSNA Brain	224 x 224	564,601	150,560	6	Yes
ISIC	224 x 224	25,979	6,088	9	No
Fundus	400 x 400	26,344	7025	5	No

Training

cd RSNA_MoCo

• Semi-Supervised Learning

  • Unsupervised training stage

    • Momentum Contrast (MoCo)

      python trainMoCo.py

    • Simple Contrastive Learning of Representations (SimCLR)

      python trainSimCLR.py

    • Momentum Contrast Version 2 (MoCoV2)

      python trainMoCoV2.py

  • Finetuning stage (Requires e.g. train10F.txt)

    • Linear Classifier which freezes all backbone encoder weights

      python MoCo_downstream.py --frozen

      or

      python MoCov2_downstream.py --frozen

    • Transfer Learning which retrains all layers (default)

      python MoCo_downstream.py

      or

      python MoCov2_downstream.py

• Supervised Learning

  • ResNet-50

    python MoCo_downstream.py --resnet

    or

    python MoCov2_downstream.py --resnet

Results

Results on RSNA data Metric: AUC

Pecentage of labeled data	MoCo	MoCo_efficient	ResNet50	MoCo_super	MoCoV2_super
100	0.9112	0.9036	0.9250	0.9419	0.9638
50	0.8967	0.8986	0.8935	0.9298	0.9519
20	0.8865	0.8958	0.8567	0.9192	0.9344
10	0.8702	0.8751	0.8476	0.9028	0.9229
5	0.8599	0.8686	0.8402	0.8837	0.8951
1	0.7751	0.7907	0.7462	0.8511	0.8177

Results on Fundus (Diabetic Retinopathy Detection) data Metric: AUC

Pecentage of labeled data	Resnet50	MoCo	MoCoV2	MoCo-FTAL
100	0.5741	0.6817	0.7390	0.7705
50	0.5541	0.6651	0.7219	0.7421
20	0.5463	0.6263	0.6607	0.6956
10	0.5432	0.5926	0.5945	0.6324
5	0.5380	0.5593	0.5944	0.5862
1	0.5369	0.5419	0.5462	0.5427

Results on ISIC data Metric: Accuracy

Pecentage of labeled data	MoCo	MoCo_V2	ResNet50	SimCLR
100	0.6467	0.6611	0.6407	0.6541
50	0.6358	0.6565	0.6565	0.6442
20	0.6200	0.6421	0.6421	0.6310
10	0.6181	0.6355	0.6355	0.6266
5	0.6131	0.6236	0.6236	0.6133
1	0.5817	0.5947	0.5947	0.5906

Resutls on Brats18 data Metric: Dice

Pecentage of labeled data	deeplabv3	deeplabv3+MoCo_super	deeplabv3+MoCo	Unet
100	0.6988	0.7261	0.7085	0.7387
50	0.6199	0.7082	0.6952	0.7110
20	0.5353	0.6255	0.6387	0.6552
10	0.3983	0.5667	0.5303	0.5545
5	0.3007	0.5267	0.5394	0.5585
1	0.1831	0.3579	0.3554	0.2655

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please contact a team member before starting work on a pull request.

License

MIT