Interactive Foundation Models to Agentic Image Analysis: Peixian Liang and Songhao Li, 01/06/26

TIA Centre Seminar Series: Dr Peixian Liang and Songhao Li Full Title: Co-evolving Computational Pathology: From Interactive Foundation Models (VISTA-PATH) to Agentic Medical Image Analysis (TissueLab) Abstract: Computational pathology is rapidly moving beyond static, single-task predictors toward interactive systems that can segment, reason, and quantify tissue morphology under expert supervision. In this talk, we present two complementary pillars of this transition from our group. First, we introduce VISTA-PATH, an interactive foundation model for histopathology image segmentation designed to resolve heterogeneous tissue structures, incorporate expert feedback, and generate pixel-level segmentations that are directly meaningful for clinical interpretation. Trained on over 1.6 million image–mask–text triplets spanning 9 organs and 93 tissue classes, VISTA-PATH outperforms existing segmentation foundation models and enables clinically relevant biomarkers, such as the Tumor Interaction Score (TIS), which shows a strong association with patient survival. Second, we presentTissueLab, an open, co-evolving agentic system (tissuelab.org) for end-to-end medical image analysis. TissueLab unifies foundation models and analytical tools within a modular tool graph, including VISTA-PATH for interactive segmentation, and translates LLM-generated plans into executable operations. Through tool-level, strategy-level, and workflow-level co-evolution, it achieves over 90% accuracy in cancer cell quantification within 10-30 minutes of expert feedback, improves lymph-node metastasis correlation from 0.827 to 0.933, and reaches a macro-AUC of 0.838 on tubule formation scoring, outperforming human-orchestrated workflows and state-of-the-art VLMs.