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

Vedang Kshirsagar, Saketh Juturu, Geetank Raipuria, Nitin Singhal

Keywords: Toxicology, Foundational Models, Multiple Instance Learning, Masked Image Modelling, Histopathology, Testes, Testicular Toxicity

TL;DR: We propose a structure-aware, self-supervised framework for learning tubule representations using MIL and show that the learned representation substantially improves automated testicular injury classification.

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, testes tubules vary widely in size and structure, making fixed-resolution patch classification ineffective. We first show that resizing tubules significantly degrades performance especially for larger sized tubules and a Multiple Instance Learning (MIL) model, offers substantial improvements. Building on this, we propose TBA-MIL, a transformer-based aggregation model with learnable positional embeddings that encodes the structure of tubules and introduce MIM-MIL, a self-supervised masked instance modelling framework that learns tubule-relevant representations from large-scale unlabeled data. Across four tubule types, TBA-MIL with MIM-MIL outperforms state-of-the-art MIL models and establishing a strong baseline for automated testicular toxicity assessment.

Primary Subject Area: Unsupervised Learning and Representation Learning

Secondary Subject Area: Application: Histopathology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 286