Task-agnostic explainable contrastive learning model for malignant cell detection in urine cytology

Abstract

Bladder cancer remains one of the most prevalent urological malignancies, where early detection is critical for improving patient outcomes. Urine cytology provides a non-invasive and cost-effective screening modality; however, the development of automated diagnostic systems is challenged by the scarcity of expert-annotated data and the limited interpretability of deep learning models. To address these challenges, we propose EnBCDet, an explainable self-supervised framework for malignant cell detection in urine cytology images. The proposed approach leverages contrastive self-supervised learning to learn discriminative cytological representations from limited annotations and incorporates a specialized backbone within a single-stage object detection architecture for efficient malignant cell localization. In addition, we introduce a novel entropy-based explanation framework that quantitatively evaluates the information content of activation maps, enabling gradient-free, class-agnostic, and task-independent interpretation of learned feature representations. Experimental evaluation on cytological samples collected from 150 individuals demonstrates that EnBCDet outperforms existing approaches, achieving a mean precision of 0.991 and a mean recall of 0.926 while maintaining computational efficiency. Clinical validation further shows over 90% agreement with expert annotations, highlighting the practical utility of the proposed system. To the best of our knowledge, this is among the first studies to systematically interpret encoder-level representations in self-supervised cytological analysis. The proposed entropy-based framework provides robust insights into backbone feature learning, advancing both diagnostic performance and explainability for computer-aided urine cytology. The analysis code supporting this study is publicly available on GitHub at https://github.com/ArathyAvinash/EnBCDet.