Boundary Contrastive Learning for Label-Efficient Medical Image Segmentation


Satoshi Kamiya (Meijo University), Kota Yamashita (Meijo University), Kazuhiro Hotta (Meijo University)
The 35th British Machine Vision Conference
PDFPoster

Abstract

Recent advances in deep learning have significantly improved the accuracy of medical image segmentation, yet the need for extensive annotations remains a challenge due to the high costs associated with practical implementation. In this study, we introduce Boundary Contrastive Learning for Label-Efficient (BCLL), a novel label-efficient learning method. The primary innovation of BCLL is the extension of contrastive loss concept used in conventional label-efficient learning methods like CLLE. We propose Boundary Contrastive Learning (BCL), which applies average pooling filters to annotated images to capture positional information about the boundary and internal regions of classes. By using features extracted from these regions, BCL computes various combinations of contrastive losses in a single image. Not only does this method bring features of the same class closer and push those of different classes apart, but it also designs contrastive losses to draw boundary region features closer to those of internal regions. This approach significantly enhances the accuracy of segmenting challenging boundary parts using only a small set of labeled data. Additionally, we have incorporated a new similarity function based on the Generalized Gaussian Distribution (GGD), named GGD-vMF, for the similarity calculations. This new similarity loss function enables enhanced learning with only minimal supervised data. Our experiments on Automatic Cardiac Diagnosis Challenge (ACDC), Synapse multi-organ segmentation (SMO), and Covid19 datasets demonstrated that BCLL achieves superior accuracy compared to the baseline and other label-efficient medical image segmentation methods. Specifically, BCLL showed an improvement in mIoU of 7.17\% with 5\% labels on ACDC, 7.74\% with 5\% labels on SMO, and 0.44\% with 10\% labels on Covid19 in comparison with baseline U-Net.

Citation

@inproceedings{Kamiya_2024_BMVC,
author    = {Satoshi Kamiya and Kota Yamashita and Kazuhiro Hotta},
title     = {Boundary Contrastive Learning for Label-Efficient Medical Image Segmentation},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0685.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection