Learning Structure-Aware Foundational Representation of Rat Testicular Tubules Using Multiple Instance Learning

Abstract

Testicular toxicity is a critical factor in preclinical drug safety assessment, yet automated modelling of testicular abnormalities remains largely unexplored. Unlike liver or kidney tissue, the testis tissue is organized into tubules that vary substantially in size and structure, making fixed-resolution patch classification ineffective. We first demonstrate that resizing tubules significantly degrades performance particularly for larger sized tubules and a Multiple Instance Learning (MIL) model offers substantial improvements. Building on this, we introduce TBA-MIL, a transformer-based aggregation model with learnable positional embeddings that encodes the structure of tubules and is pre-trained using a self-supervised Masked Instance Modelling (MIM-MIL) framework, learning tubule representations from large-scale unlabeled data. Across four tubule types, TBA-MIL with MIM-MIL outperforms state-of-the-art MIL models and establishes a strong baseline for automated testicular toxicity assessment. Additionally, we evaluate the proposed framework on an independent toxicological study and show that the predicted abnormality distributions significantly differentiate control and treated animal tissues, consistent with expert pathologists’ assessment.