acceptodds
Withdrawn from ICLR 2027

PlanPath: Exploration-Guided Orchestration for Patient-Level Pathology Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A patient's pathology case spans tens of slides from distinct specimens and anatomical sites, each carrying non-redundant, site-specific evidence. Forming a patient-level conclusion therefore requires locating the slides relevant to a given question and integrating complementary findings across them. However, existing pathology agents analyze one slide at a time and describe morphology rather than measure it, leaving evidence acquisition across a patient's specimens and patient-level quantification underexplored. To address this gap, we propose PlanPath, an agentic framework for patient-level pathology visual question answering (VQA). PlanPath turns the question into an exploration plan that specifies the evidence required, routes to the slides likely to carry it, and drives coarse-to-fine acquisition from tissue architecture down to individual cells. Tissue- and cell-level tools supply the counts, proportions, and extents, which PlanPath aggregates across slides into a traceable answer. To enable evaluation, we build two patient-level VQA datasets for renal and gastric cancer, averaging 15 WSIs per patient, with pathologist-verified question–answer pairs derived from diagnostic reports. Experimental results show that PlanPath outperforms the strongest baseline by 9.4% and 12.1% in overall accuracy on the two datasets and further improves existing slide-level agents when plugged into our framework. The dataset and code are available at: https://anonymous.4open.science/r/PlanPath.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational Pathology, Patient-Level Visual Question Answering, Pathology Agents

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_f8dLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PlanPath addresses patient-level pathology visual question answering (VQA), answering questions from a patient’s full collection of whole-slide images (WSIs) rather than isolated slides. Its core workflow integrates several specialized components: an Agent Brain (Qwen3.5-4B) translates queries into exploration plans, a TITAN-based Slide Router ranks candidate slides via text–slide similarity, and PLIP performs ROI relevance ranking. Multi-magnification region examinations are further supported by Patho-R1-7B for morphological descriptions, alongside tissue-prototype analysis and HistoPLUS-based cell quantification for spatial or numerical evidence. By tracking slide, ROI, and magnification provenance, the system iteratively integrates evidence before generating a final answer. To evaluate PlanPath, the authors introduced gastric and renal datasets spanning 213 patients, 3,245 WSIs, and 1,623 expert-reviewed question–answer pairs. PlanPath achieved 60.24% accuracy on PatientVQA-Gastric and 72.71% on PatientVQA-Renal, outperforming the strongest baselines by 9.43 and 12.14 percentage points, respectively. Additional experiments cover agent compatibility, component ablations, quantitative evaluation, and qualitative cases. However, whether the evaluation rigorously isolates visually grounded, cross-slide reasoning and validates specific efficiency claims remains an open question.

Critical strengths
  1. Meaningful Problem and Principled System Architecture:

The paper addresses a challenging, understudied problem by transitioning from isolated-slide VQA to patient-level synthesis across multiple WSIs. PlanPath offers a coherent, provenance-aware architecture combining slide routing, multi-scale ROI exploration, and specialized tools. Crucially, its explicit exploration plans and evidence gates enhance inspectability over opaque end-to-end models.

  1. Extensive Empirical Validation and Strong Results:

The study provides a broad evaluation across two novel datasets (213 patients, 3,245 WSIs, 1,623 QA pairs), demonstrating notable accuracy gains over strong baselines (+9.43% on Gastric, +12.14% on Renal). Beyond headline metrics, comprehensive ablations, compatibility tests, and qualitative cases effectively illuminate the system's routing and evidence integration mechanisms.

Critical weaknesses
  1. Unclear Evaluation of True Cross-Slide Reasoning and Visual Grounding:

While the paper emphasizes transitioning to patient-level VQA, the benchmark evaluation fails to isolate genuine cross-slide reasoning. Because pathology report-derived QA pairs may be answerable from a single informative WSI, local visual recognition, or clinical priors, it remains unclear how many questions strictly require multi-slide synthesis. The pathologist review protocol lacks crucial details (e.g., annotation guidelines, inter-annotator agreement, and explicit counts of multi-slide dependency), and performance is not disaggregated by single- vs. multi-slide sufficiency. Without separating these cases, the reported gains cannot convincingly prove that the improvement stems from cross-slide integration rather than single-slide recognition.

  1. Insufficient Evidence for Quantitative Claims, Statistical Rigor, and True Efficiency:

The validation of quantitative reasoning is constrained by an unconfirmed, small sample size (appearing to be only 18 questions based on step increments), where tolerance metrics (±5\pm 5 to ±20\pm 20) lack clinical justification and tools like cell quantification yield a marginal ablation impact (+0.43%). Furthermore, accuracy metrics are presented solely as point estimates without confidence intervals or significance testing to account for patient-level clustering. Finally, the efficiency analysis relies strictly on prompt-token counts, omitting critical end-to-end operational costs such as tool latency, GPU time, generated output tokens, and monetary expense.

Questions for authors

Q1. Cross-Slide Dependency and Evaluation Protocol: Could the authors clarify what proportion of evaluation questions genuinely require cross-slide or multi-site synthesis, as opposed to single-slide sufficiency? Please provide the specific pathologist screening protocol, inter-annotator agreement, and per-question evidence mappings (e.g., required WSIs/ROIs). Reporting separate accuracy for single- vs. multi-slide questions is essential to validate the core task claims.

Q2. Quantitative Sub-benchmark and Tool Grounding: Please confirm the exact sample size for the quantitative subset (the 5.56% increments suggest n=18n=18) and specify how ground-truth values were derived. Were tissue-prototype and HistoPLUS outputs validated against expert annotations? Additionally, what are the clinical rationales and units for the ±5\pm 5 to ±20\pm 20 tolerance thresholds?

Q3. Statistical Rigor and Patient-Level Uncertainty: Could the authors provide patient-level resampled confidence intervals (e.g., paired bootstrap) or significance tests for the primary accuracy results? Since multiple questions originate from the same patient, treating QA pairs as independent observations may underestimate uncertainty and obscure whether the headline gains (+9.43% and +12.14%) are statistically significant.

Q4. End-to-End Efficiency and Implementation Details: To provide a comprehensive efficiency profile, could the authors report wall-clock latency, tool-call frequency, output token counts, GPU runtime, and monetary costs alongside prompt tokens? Please also clarify the "Slide-T" setting, exact model/API decoding parameters, and whether any WSIs in the tissue prototype banks overlap with the evaluation cohorts.

Reviewer_ywDCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper the authors proposed an agentic framework for integrating visual evidence from multiple slides for pathological reasoning at the patient level. The work appears to be technically sound. But I also have a few concerns, especially regarding regarding how evidence is integrated and decisions are made. Experimental validation of the proposed framework is extensive.

Critical strengths
  1. Overall the paper is well written. But there are a few places that require clarification.

  2. The idea of integrating evidence from multiple slides is reasonable and generally supported by the experimental results.

Critical weaknesses
  1. Integration of evidence across scale is important. The paper mentioned an evidence chain (page 6 line 290). But the paper is unclear on how the evidence chair is constructed. I could not find an answer to this question in the Supplementary Materials either.

  2. Integration of evidence across different slides is also critical. The paper mentioned an evidence gating scheme. But how are the thresholds set? How sensitive is the final answer to threshold settings?

Questions for authors
  1. Can you provide a detailed description of how evidence is aggregated across scale? How is this implemented? By Qwen3.5-4B?

  2. Can you provide a detailed description of how evidence is integrated across different slides? How is this implemented? By Qwen3.5-4B?

  3. In real world cases, evidences often are probabilistic, carry uncertainty and are contradictory in some cases. How are the uncertainty and contradiction handled in general in this study? By Qwen3.5-4B?

  4. What are the limitations of this study?

Reviewer_4HjtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces an exploration-guided framework (PlanPath) for patient-level pathology VQA. Unlike conventional slide-level methods, PlanPath aims to answer questions by integrating visual evidence across multiple whole-slide images.

PlanPath consists of question-driven exploration planning, slide routing, hierarchical tissue- and cell-level analysis, iterative evidence acquisition, and cross-slide evidence integration. The authors also introduce two patient-level VQA datasets, PatientVQA-Gastric and PatientVQA-Renal, and demonstrate improved performance over existing supervised-based models and agent-based models.

Critical strengths
  1. Well-motivated problem formulation.
  • The transition from slide-level to patient-level pathology reasoning is clinically meaningful. Many diagnostic questions may require integrating complementary evidence across multiple WSIs, and the paper addresses an important limitation of existing approaches.
  1. New patient-level benchmarks.
  • The construction of two pathology VQA datasets is a valuable contribution. The involvement of pathologists in reviewing question–answer pairs helps establish the clinical relevance of the evaluation.
  1. Clear framework design.
  • The overall architecture is clearly presented, and the major components and their interactions are relatively easy to understand.
Critical weaknesses
  1. Unclear necessity of multi-slide reasoning across the proposed benchmarks. A central motivation of the paper is that patient-level questions require evidence integration across multiple WSIs. However, the paper does not provide a systematic analysis of how many questions genuinely require multiple slides to answer. It would be useful to categorize the questions according to their evidence requirements, such as:
  • Questions answerable from a single WSI.
  • Questions requiring evidence from multiple WSIs.
  • Questions requiring exhaustive examination of the available slides.

Evaluating performance separately on these subsets would provide stronger evidence for the necessity and effectiveness of the proposed patient-level framework.

  1. Insufficient analysis of the contributions of the proposed components and mechanisms
  • Although the paper provides ablation studies in Table 4, it is still unclear what contributes most to the performance improvement. The paper mainly reports final accuracy without evaluating intermediate results, such as the quality of selected slides, ROIs, and collected evidence. It is difficult to determine whether the performance improvements come from the proposed exploration strategy or simply from combining existing pathology models.
  • For example, when PlanPath produces an incorrect answer, is it because the Slide Router fails to retrieve the relevant slides, the Tissue Analysis Tool selects incorrect regions, or the Cell-level Quantification Tool produces inaccurate measurements?
  1. Limited evaluation of generalization and the effectiveness of slide routing
  • The experiments are conducted on two newly constructed patient-level datasets, both from specific pathology domains. Additional evaluation on established public WSI benchmarks (datasets from WSI-LLava or SlideChat) would help assess the generalizability of the framework. For example, if we already know the ground-truth slide selection in a slide-level VQA task, would PlanPath perform better than the slide-level method?
  1. Lack of evaluation of intermediate reasoning reliability

PlanPath relies on multiple model-driven steps, including exploration planning, slide and ROI selection, tissue-level analysis, morphological interpretation, and evidence integration. However, the reliability of these intermediate steps is not sufficiently evaluated. For instance, an incorrect exploration plan or an inaccurate tissue-level prediction could influence subsequent navigation decisions and ultimately affect the final answer. The paper does not systematically investigate such error propagation or determine which stages are most responsible for unsuccessful predictions. An evaluation of intermediate evidence quality, potentially supported by expert annotations, would substantially strengthen the paper's claims regarding evidence-guided reasoning.

  1. Insufficient validation of the Evidence Gate and stopping criteria

The Evidence Gate is a central component of the proposed adaptive exploration framework. Although the paper describes conceptual criteria based on evidence coverage, quality, target satisfaction, and spatial consistency, the practical behavior and reliability of these stopping decisions remain insufficiently validated.

For example:

  • How are the stopping thresholds determined?
  • How often does exploration terminate because sufficient evidence has been collected vs. because the analysis budget is exhausted?
  • How frequently does the system stop prematurely?
  • How sensitive is the final performance to different stopping thresholds? The reported prompt-token efficiency provides indirect evidence of reduced computation, but does not establish whether the adaptive stopping mechanism makes reliable decisions. Comparisons against fixed exploration budgets and analyses of exploration trajectories would be particularly valuable.
  1. Limited expert validation and representativeness of qualitative case studies

Although pathologists participate in validating the constructed question–answer pairs, the paper does not provide a systematic expert evaluation of the intermediate evidence and reasoning trajectories produced by PlanPath.

The presented case studies illustrate successful reasoning processes, but it remains unclear how representative these examples are of the overall system behavior.

For example, the quantitative case study would benefit from expert verification of the selected regions, tissue-level measurements, and cross-slide evidence aggregation. To what extent do domain experts agree with the decisions and results at each step?

Questions for authors

What is the main bottleneck limiting the performance of PlanPath? Could the authors provide an error analysis to identify which stage contributes most to incorrect predictions?

Reviewer_efqcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper propose their agentic workflow PlanPath with novel patient-level pathology datasets. PlanPath answers pathology questions from a patient's collection of whole-slide images. It plans evidence needs, ranks slides, explores regions at several scales, and combines tissue, cell, and morphology tools. The paper also introduces two expert-reviewed patient-level VQA datasets. The system improves overall accuracy under the reported comparisons.

Critical strengths
  1. The task and datasets are useful. The study provides question and corresponding sildes from real patients reviewed by two pathologists. This will make real contribution towards community.
  2. The system-level gains are substantial. Total accuracy reaches 60.24% and 72.71%. Gains remain under both baseline aggregation strategies, and adding PlanPath improves the overall results for existing agents.
Critical weaknesses
  1. The proposed system lacks in-parameter learning or further proofs to show potential extrapolation. Although the paper proposes a useful agentic workflow and it improves patient-level pathology tasks, we do not know whether this capability can be learned or extrapolated to more patient-level tasks.
  2. It is unclear whether the improvement is due to the tools or the claimed agentic workflow design. For open questions, the baseline aggregator must select one slide-level answer verbatim, while PlanPath can combine evidence and generate a new answer. A shared-tool baseline with unrestricted evidence fusion is needed to attribute the gains more clearly.
  3. The efficiency statistics using only prompt token numbers are oversimplified.
  4. The current data scale is not fully convincing to support its claimed patient-level advantage, since the datasets only cover two diseases and related tools in these two scenarios.
Questions for authors
  1. Could you clarify whether the performance gains are beneficial from your workflow design?
  2. Could you provide results on other data sources corresponding to different diseases with different tools?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.