Latent Visual Diffusion Reasoning for Explainable Pathology Grading

Faris H. Rizk, Nan Xi, Junsong Yuan

Keywords: Latent Visual Diffusion Reasoning; Explainable Pathology Grading

Abstract:

Pathology grading (e.g., tumor severity assessment) is a fundamental task in clinical workflows and plays a critical role in guiding diagnosis, treatment planning, and prognosis. Achieving accurate grading, however, requires extensive domain expertise and a structured, multi-step reasoning process that integrates subtle visual cues across the imaging volume. Most existing computational approaches simplify this task into a direct regression or classification problem, predicting only the final grade without revealing the intermediate reasoning steps that lead to the decision. This lack of transparency limits their utility in clinical settings, where interpretability and alignment with human reasoning are essential.

To address these limitations, we propose Latent Visual Diffusion Reasoning (LVDR), a framework that explicitly models the visual reasoning process in a latent reasoning space. LVDR formulates pathology reasoning as a diffusion process that evolves sequentially along the temporal axis of the scan: the first slice is treated as a noisy initialization, and the model progressively denoises through subsequent slices to produce a coherent reasoning trajectory. This formulation enables LVDR to learn step-by-step, interpretable reasoning paths that uncover how the model integrates visual evidence across slices to reach the final grading decision.

Extensive experiments on three pathology grading benchmarks demonstrate that LVDR achieves performance on par with state-of-the-art methods while providing substantially more interpretable visual reasoning trajectories. These trajectories illuminate the internal decision-making process and offer a level of transparency that is critical for clinical adoption.

Primary Subject Area: Interpretability and Explainable AI

Secondary Subject Area: Detection and Diagnosis

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 112