xPath: Human-AI Diagnosis in Pathology with Multi-Criteria Analyses and Explanation by Hierarchically Traceable Evidence

Data-driven AI promises support for pathologists to discover sparse tumor patterns in high-resolution histological images. However, from a pathologist's point of view, existing AI suffers from three limitations: (i) a lack of comprehensiveness where most AI algorithms only rely on a single criterion; (ii) a lack of explainability where AI models tend to work as 'black boxes' with little transparency; and (iii) a lack of integrability where it is unclear how AI can become part of pathologists' existing workflow. Based on a formative study with pathologists, we propose two designs for a human-AI collaborative tool: (i) presenting joint analyses of multiple criteria at the top level while (ii) revealing hierarchically traceable evidence on-demand to explain each criterion. We instantiate such designs in xPath -- a brain tumor grading tool where a pathologist can follow a top-down workflow to oversee AI's findings. We conducted a technical evaluation and work sessions with twelve medical professionals in pathology across three medical centers. We report quantitative and qualitative feedback, discuss recurring themes on how our participants interacted with xPath, and provide initial insights for future physician-AI collaborative tools.