SOoD: Self-Supervised Out-of-Distribution Detection Under Domain Shift

This repository contains the Pytorch implementation of the proposed method SOoD: Self-Supervised Out-of-Distribution Detection Under Domain Shift for Multi-Class Colorectal Cancer Tissue Types ( written by Guillaume Vray (https://github.com/GuillaumeVray/SOoD)), which has been recently accepted at ICCV 2021 Workshop (CVAMD 2021).

A) Dependencies

Install dependencies via executing the setup.sh script using the conda environment. We use Python 3.9 and Pytorch 1.9.

B) Training the SOoD Model

To visualize the training process, setup the visdom server by running:

visdom -p 1076 -env_path path_to_repo/src/Visdom

To train our model, run:

python main.py SOoD Kather expname --aug_type E/H --n_prototypes 16 --ssl_mode sinkhorn

Run unsupervised fine-tuning with:

python main.py SOoDftUns Kather expname2 --ssl_ckpt Experiments/expname/Checkpoints/SOoD_best_loss_source.pth --aug_type E/H --n_prototypes 16 --ssl_mode sinkhorn

Run supervised fine-tuning with:

python main.py SOoDftClassifier Kather expname3 --ssl_ckpt Experiments/expname/Checkpoints/SOoD_best_loss_source.pth --aug_type E/H --n_prototypes 16 --ssl_mode sinkhorn --pct_data percentage_data

C) Set the parameters to run baselines and ablations:

• 3heavy+1style: --lmd_h 3 --lmd_ood 1

• 1heavy+3style: --lmd_h 1 --lmd_ood 3

• 0heavy+1style: --lmd_h 0 --lmd_ood 1

• 1heavy+0style: --lmd_h 1 --lmd_ood 0

• 8 prototypes: --n_prototypes 8

• 24 prototypes: --n_prototypes 24

• colorization: --aug_type E/H/S

• Swav Baseline: --aug_type normal --color_transformation --ssl_mode sinkhorn

• DINO Baseline: --aug_type normal --color_transformation --ssl_mode S/T

• Classifier: --aug_type normal --ssl_mode no

D) Testing

To test any model with name model_name, simply run:

python main.py model_name Kather expname --checkpoint path/to/checkpoint --test 

E) Our results and checkpoints

Unsupervised methods

Method	AUROC	AUPRC	Checkpoint
SOoD-pretrain	88.38	80.43	ckpt
SOoD-finetune Unsupervised	92.77 +/- 0.48	90.90 +/- 1.00	ckpt
SOoD-pretrain 0heavy	84.63	75.76	ckpt
SOoD-finetune 0heavy	91.60 +/- 0.59	90.62 +/- 0.54	ckpt
SOoD-pretrain 3heavy	85.18	76.41	ckpt
SOoD-finetune 3heavy	90.77 +/- 0.47	87.95 +/- 0.33	ckpt
SOoD-pretrain 0style	83.95	75.63	ckpt
SOoD-finetune 0style	85.87 +/- 3.25	84.56 +/- 2.26	ckpt
SOoD-pretrain 3style	82.76	71.80	ckpt
SOoD-finetune 3style	90.99 +/- 0.28	88.64 +/- 0.39	ckpt
SOoD-pretrain 8prots	82.52	71.44	ckpt
SOoD-finetune 8prots	88.85 +/- 0.57	85.30 +/- 0.22	ckpt
SOoD-pretrain 24prots	86.70	79.60	ckpt
SOoD-finetune 24prots	89.24 +/- 0.40	84.71 +/- 0.43	ckpt
SOoD-pretrain ColorStyle	82.44	75.23	ckpt
SOoD-finetune ColorStyle	88.58 +/- 0.73	86.54 +/- 0.57	ckpt

Linear classification baselines on frozen features

Method	Linear	KNN	Checkpoint
SOoD-finetune Supervised 100%	73.24 +/- 0.39	83.45	ckpt
SOoD-finetune Supervised 20%	73.39 +/- 0.69	-	ckpt
SOoD-finetune Supervised 10%	73.24 +/- 0.82	-	ckpt
SOoD-finetune Supervised 1%	62.59 +/- 1.42	-	ckpt
SwAV 100%	48.83 +/- 1.83	41.72	ckpt
DINO 100%	42.03 +/- 8.51	32.20	ckpt
Supervised Source 100%	65.13 +/- 3.57	41.50 +/- 3.84	ckpt
Supervised Translated 100%	78.31 +/- 5.98	77.48 +/- 1.61	ckpt

---

Citation

If you find this work useful, please cite the paper:

@inproceedings{bozorgtabar2021sood,
  title={SOoD: Self-Supervised Out-of-Distribution Detection Under Domain Shift for Multi-Class Colorectal Cancer Tissue Types},
  author={Bozorgtabar, Behzad and Vray, Guillaume and Mahapatra, Dwarikanath and Thiran, Jean-Philippe},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3324--3333},
  year={2021}
}

---

---

Licence

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.