TEAR: Semi-Supervised Medical Image Classification with Temporal Knowledge-Aware Regularization

This repository is an official PyTorch implementation of the paper "Semi-Supervised Medical Image Classification with Temporal Knowledge-Aware Regularization" from MICCAI 2022.

Introduction

In this work, we propose TEmporal knowledge-Aware Regularization (TEAR) for semi-supervised medical image classification. Instead of using hard pseudo labels to train models roughly, we design Adaptive Pseudo Labeling (AdaPL), a mild learning strategy that relaxes hard pseudo labels to soft-form ones and provides a cautious training. AdaPL is built on a novel theoretically derived loss estimator, which approximates the loss of unlabeled samples according to the temporal information across training iterations, to adaptively relax pseudo labels. To release the excessive dependency of biased pseudo labels, we take advantage of the temporal knowledge and propose Iterative Prototype Harmonizing (IPH) to encourage the model to learn discriminative representations in an unsupervised manner. The core principle of IPH is to maintain the harmonization of clustered prototypes across different iteration.

Dependencies

All experiments use PyTorch library. We recommend installing following package versions:

•    python==3.7

•    pytorch==1.6.0

•    MedPy==0.4.0

•    scipy==1.5.4

Dependency packages can be installed using following command:

pip install -r requirements.txt

Quickstart

Training

python Train_TEAR.py 
    --gpu=0 
    --dataset=isic2018 
    --n-classes=7 
    --backbone=alexnet 
    --n-labeled=350 
    --adapl=True
    --batchsize=128 
    --thr=0.95 
    --n-epoches=256 
    --setting=ISIC_350

Bonus

Application on natural images

The strategy AdaPL can also bring performance improvements on many natural image datasets, including CIFAR-10, CIFAR-100, SVHN, etc, especially in low label regimes. To have a try, you simply have to add the dataloaders in ./datasets/cifar.py, and turn on the switch of AdaPL: --adapl=True. Enjoy playing on your own datasets :-D

Citation

@inproceedings{yang2022d2,
  title={Semi-Supervised Medical Image Classification with Temporal Knowledge-Aware Regularization},
  author={Qiushi Yang, Xinyu Liu, Zhen Chen and Yixuan Yuan},
  booktitle= {MICCAI},
  year={2022}
}

Acknowledgements

• The implementation of baseline method is based on: CoMatch and FixMatch.