Title: Emulating Pathological Expertise: Tissue-Intelligent Reasoning for Whole Slide Images

Abstract:

The field of computational pathology has seen swift progress, fueled by the development of specialized image encoders and a rising enthusiasm for employing vision-language models to address natural-language inquiries regarding diseases. However, the fundamental challenge of pathology question-answering persists, primarily because a single gigapixel slide holds vastly more data than is typically required to answer a specific query. While human pathologists manage the intricacies of tissue and morphology by conducting broad scans and then zooming in selectively based on clinical needs, current computational models often fall short. These existing approaches typically depend on uniform patch sampling or wide-ranging attention maps, which frequently result in equal focus on irrelevant areas while missing crucial visual evidence.

To bridge this gap, this study aims to align model behavior with human slide examination techniques. We introduce HistoSelect, a retrieval framework that is question-guided, tissue-aware, and operates on a coarse-to-fine basis. This system comprises two primary elements: a group sampler designed to pinpoint tissue regions relevant to the query, followed by a patch selector that extracts the most significant patches from those areas. By focusing exclusively on the most informative patches, the proposed method achieves substantial efficiency gains, cutting down visual token consumption by an average of 70%. Simultaneously, it enhances accuracy across three distinct pathology QA benchmarks.

Assessed against a dataset of 356,000 question-answer pairs, our strategy surpasses current state-of-the-art methods and generates responses anchored in interpretable regions that align with pathologist standards. These findings indicate that integrating human-like search strategies and attention mechanisms into Whole Slide Image (WSI) reasoning represents a viable path toward developing practical and dependable pathology Vision-Language Models (VLMs). The source code for this project is accessible at https://github.com/winston52/HistoSelect.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC