PathFound: An Agentic Multimodal Model Activating Evidence-seeking Pathological Diagnosis

TL;DR

PathFound addresses the gap between static, read-once pathological inference and dynamic clinical diagnostics by introducing an agentic, evidence-seeking framework. It combines a Slide Highlighter, Vision Interpreter, and Diagnostic Reasoner within a 3-stage 3E protocol (Exploration, Execution, Exploitation) to progressively formulate hypotheses, obtain targeted evidence, and finalize diagnoses. Across RCC subtyping and nuclear grading, PRAD Gleason grading, and pan-cancer invasion detection, PathFound achieves state-of-the-art or near-top performance, with high proactive evidence mentioning (PEMR) in fine-grained tasks. The approach demonstrates the value of iterative, hypothesis-driven reasoning and targeted information retrieval for improving diagnostic accuracy and interpretability in computational pathology.

Abstract

Recent pathological foundation models have substantially advanced visual representation learning and multimodal interaction. However, most models still rely on a static inference paradigm in which whole-slide images are processed once to produce predictions, without reassessment or targeted evidence acquisition under ambiguous diagnoses. This contrasts with clinical diagnostic workflows that refine hypotheses through repeated slide observations and further examination requests. We propose PathFound, an agentic multimodal model designed to support evidence-seeking inference in pathological diagnosis. PathFound integrates the power of pathological visual foundation models, vision-language models, and reasoning models trained with reinforcement learning to perform proactive information acquisition and diagnosis refinement by progressing through the initial diagnosis, evidence-seeking, and final decision stages. Across several large multimodal models, adopting this strategy consistently improves diagnostic accuracy, indicating the effectiveness of evidence-seeking workflows in computational pathology. Among these models, PathFound achieves state-of-the-art diagnostic performance across diverse clinical scenarios and demonstrates strong potential to discover subtle details, such as nuclear features and local invasions.

Figures (13)

• Figure 1: Comparison between conventional one-pass diagnosis and PathFound’s evidence-seeking diagnostic paradigm. (A) One-pass diagnosis: existing multimodal models analyze a whole-slide image once and generate a final prediction without revisiting the slide or refining the diagnostic hypothesis. (B) Evidence-seeking diagnosis with PathFound: the diagnosis proceeds through an outer loop of hypothesis formulation, targeted evidence acquisition (via slide re-observation or external tests), and conclusion refinement, enabling progressive, uncertainty-aware diagnostic reasoning.
• Figure 2: A complete diagnostic path found by PathFound. The path starts with an initial diagnosis stage, where it generates a list of possible diagnoses and plans the next steps based on preliminary information. Then it transitions into an evidence-seeking mode, re-observing the slides with specific purposes and collecting external examination results. It ends with a final decision stage, in which a precise diagnosis is derived with the additional evidence. Notably, there may be more routes for practical use.
• Figure 3: A detailed view of the slide highlighter.
• Figure 4: Qualitative Results on Nuclear Grading from TCGA-RCC.
• Figure 5: Ablation study on the combination of input RoIs when screening RCC.