WeakSupCon: Weakly Supervised Contrastive Learning for Encoder Pre-training
Abstract
Weakly supervised multiple instance learning (MIL) is a challenging task given that only bag-level labels are provided, while each bag typically contains multiple instances. This topic has been extensively studied in histopathological image analysis, where labels are usually available only at the whole slide image (WSI) level, while each WSI could be divided into thousands of small image patches for training. The dominant MIL approaches focus on feature aggregation and take fixed patch features as inputs. However, weakly supervised feature representation learning in MIL settings is always neglected. Those features used to be generated by self-supervised learning methods that do not utilize weak labels, or by foundation encoders pre-trained on other large datasets. In this paper, we propose a novel weakly supervised feature representation learning method called Weakly Supervised Contrastive Learning (WeakSupCon) that utilizes bag-level labels. In our method, we employ multi-task learning and define distinct contrastive losses for samples with different bag labels. Our experiments demonstrate that the features generated using WeakSupCon with limited computing resources significantly enhance MIL classification performance compared to self-supervised approaches across three datasets. Our WeakSupCon code is available at github.com/BzhangURU/Paper_WeakSupCon
Cite
@article{arxiv.2503.04165,
title = {WeakSupCon: Weakly Supervised Contrastive Learning for Encoder Pre-training},
author = {Bodong Zhang and Hamid Manoochehri and Xiwen Li and Beatrice S. Knudsen and Tolga Tasdizen},
journal= {arXiv preprint arXiv:2503.04165},
year = {2025}
}
Comments
Medical Image Computing and Computer Assisted Intervention (MICCAI) 2025 workshop on Efficient Medical AI