Pathology agents stop answering one slide at a time: PlanPath turns a question into an exploration plan and reaches 72.71% total accuracy on renal cancer
Synopsis
A pathology case often spans dozens of slides from different specimens, yet existing pathology agents analyze one slide at a time and describe morphology rather than measure it. PlanPath turns a diagnostic question into an exploration plan, routes to the slides likely to carry the evidence, then zooms from tissue architecture down to cells for counts and proportions, aggregating cross-slide evidence into a traceable answer. On patient-level renal and gastric cancer datasets, PlanPath reaches total accuracy of 72.71% and 60.24%, 12.14 and 9.43 percentage points above the strongest baseline.
Interpretation
The unit of pathology question answering moves from a single slide to a patient's whole slide collection, where the answer must gather complementary evidence across slides and scales. Existing pathology agents analyzed one slide at a time and described morphology rather than measuring it, leaving evidence acquisition across specimens and patient-level quantification largely unexplored. On the patient-level renal and gastric cancer datasets, PlanPath reaches total accuracy of 72.71% and 60.24%, 12.14 and 9.43 percentage points above the strongest baseline.
The system first writes the question into an exploration plan that specifies what evidence is needed, at which scale to acquire it, and when to stop, then uses that plan to decide which slides to examine. Slide routing uses TITAN to encode slides and text descriptions into one semantic space and rank by similarity; routing scores only set exploration order and are not treated as diagnostic evidence. The two datasets contain 213 patients, 3,245 whole-slide images and 1,623 pathologist-verified question-answer pairs, averaging 21.41 slides per patient in gastric cancer and 9.34 in renal cancer.
Evidence acquisition runs coarse to fine: tissue-level structure and distribution come first, then selected high-magnification regions are passed to a cell quantification tool for counts, proportions and densities. The tissue-level tool builds a soft tissue map from a tissue feature prototype bank, the cell-level tool uses HistoPLUS, and morphological descriptions come from Patho-R1-7B. Removing the tissue-level analysis tool lowers overall accuracy from 65.62% to 62.05%, the largest drop among the ablation settings.
Perspective
The demonstrated scope is two patient-level visual question answering datasets for renal and gastric cancer, whose questions and reference answers come from diagnostic reports and were cross-reviewed by two pathologists; pathology reports are not provided to models during inference, and reference answers are reserved for evaluation. Component-swapping experiments cover slide-level agents including PathAgent, PathNavigate, GIANT and BEACON.
Ablations show component contributions vary by question type and are not strictly additive, with cell-level quantification benefits more dependent on the specific question. In compatibility experiments, open-ended gains differ by underlying agent, and GIANT shows small decreases on that question type. When exploration ends before evidence requirements are met, the answer is marked low-confidence and evidence-insufficient.