TSMCR: Two-stage Supervised Multi-modality Contrastive Representation for Ultrasound-based Breast Cancer Diagnosis
Keywords: multimodality breast ultrasound; supervised contrastive learning; representation learning; deep support vector machine
Verify Author List: I have double-checked the author list and understand that additions and removals will not be allowed after the submission deadline.
Abstract: Contrastive learning has demonstrated great performance in breast cancer diagnosis. However, few existing works inspect label information in contrastive representation learning, especially for multi-modality ultrasound scenes. In this work, a two-stage supervised multi-modality contrastive representation classification network (TSMCR) is proposed for assisting breast cancer diagnosis on the multimodality ultrasound. TSMCR consists of two-stage supervised multimodality contrastive learning (SMCL) and deep support vector machine (DSVM). By a novel contrastive loss, SMCL handles the consistency between modalities and the sample separability. Further, two-stage SMCL learns expressive representation by gradually pulling the similar samples of positive pairs closer and pushing the dissimilar samples of negative pairs apart in the projection space. Besides, on the fusion of the multi-level contrastive representation, DSVM is to jointly learn the representation network and classifier again in a unified framework to improve the generation performance. The experimental results on the multimodality ultrasound dataset show the proposed TSMCR achieves superior performance with an accuracy of 87.51%, sensitivity of 86.67%, and specificity of 88.36%.
A Signed Permission To Publish Form In Pdf: pdf
Url Link To Your Supplementary Code: https://github.com/351gbm/TSMCR-ACML
Primary Area: Applications (bioinformatics, biomedical informatics, climate science, collaborative filtering, computer vision, healthcare, human activity recognition, information retrieval, natural language processing, social networks, etc.)
Paper Checklist Guidelines: I certify that all co-authors of this work have read and commit to adhering to the guidelines in Call for Papers.
Student Author: Yes
Submission Number: 308
Loading