Contrastive Learning for Limited Medical Data – A Bipartite Strategy for Detection Tasks

Training deep learning models for medical applications often requires manual annotation of data at scale, which can become impossible under circumstances as annotated medical data is often scarce and limited. Yet, availability of abundant unlabelled data can be exploited for the development of better deep models. The proposed approach leverages a bipartite strategy, where a contrastive pre-training method is employed to enhance the representation of the backbone network, followed by fine-tuning with limited labeled data. Experiments on several medical detection tasks were conducted. Results showed significant performance improvements, with our method achieving 45.67% mAP using only 50% of the LIVECell dataset, compared to the benchmark’s 48.43% with the full dataset. With merely 2% of the dataset, our strategy also outperformed training from scratch methods by 5.22%. To test the universal applicability of the proposed strategy, we applied it to another limited dataset, VinBig, with results confirming benefits of the strategy. These findings highlight the efficacy of self-supervised learning for enhancing detection performance with minimal annotated data.