Abstract
Large language model (LLM)-based agents have shown potential
for whole-slide image (WSI) analysis by approximating pathologists’
diagnostic workflow: navigating across gigapixel slides, zooming
into suspicious regions, and collecting visual evidence before making diagnostic
claims. However, existing evaluations focus mainly on final accuracy,
leaving unclear where the agent looked, which regions supported
each claim, and whether the claimed evidence was ever inspected. Without
such traces, agent behaviour is difficult to replay, compare, and audit
across models. We present PathTrace, a trace-based evidence harness
and offline auditing interface for WSI pathology agents. PathTrace converts
each agent run into a standardized trace that records percentagecoordinate
views, marked evidence, and diagnostic claims. PathTrace
also provides a trace-auditing interface for replaying navigation paths,
inspecting the exact patches seen by the model, and flagging missed tumour
regions, off-tumour evidence, and trace-inconsistent citations. We
conduct a multi-model study on 250 PANDA prostate biopsies. Path-
Trace reveals a failure mode that standard accuracy hides: one open agent
achieves 0.83 binary accuracy for cancer detection, yet 77% of its evidence
pins fall outside annotated tumour regions. We additionally demonstrate
a simple audit-informed grounding check that improves grade agreement
from κ=0.06 to κ=0.41 without retraining. PathTrace therefore provides
a practical framework for inspecting evidence provenance beyond final
diagnosis accuracy.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_034.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{LimChi_PathTrace_MICCAISAT2026,
author = { Lim, Chiew Hui},
title = { { PathTrace: A Trace-Based Evidence Harness for Auditing Whole-Slide Pathology Agents } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}