bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / pathology / cpathagent
PathologyImaging
Westlake UniversityZhejiang UniversityThe Ohio State UniversityReleased May 2025

CPathAgent

Agent-based pathology model that navigates whole-slide images by zooming and panning like a pathologist, scoring 88.6% on the PathMMU-HR2 benchmark.

17Openness

Where to run it

No providers recorded yet. Browse all providers

CPathAgentPathologyWestlake University

Computational pathology models generally take one of two shapes: a general-purpose patch encoder feeding a multiple-instance-learning head that classifies a whole slide, or a multimodal model that emits a report straight from image input. Both skip the step that defines actual diagnostic practice. A pathologist begins at low magnification to survey the slide, picks out suspicious territory, zooms in progressively, and assembles a diagnosis from observations gathered across magnifications — a process whose intermediate reasoning is exactly what existing models discard.

CPathAgent, from Westlake University with collaborators at Zhejiang University and The Ohio State University, reframes whole-slide analysis as agentic navigation. The model decides where to look and at what magnification, issuing zoom and move actions over a gigapixel slide, and narrates its findings as it goes. The output is a diagnostic summary accompanied by the viewing path that produced it, which makes the reasoning auditable rather than opaque.

A second contribution is PathMMU-HR2, an expert-validated benchmark for "huge region" analysis — the intermediate scale between a single patch and an entire slide that most closely matches how pathologists actually work, and which no prior benchmark covered. CPathAgent was presented as a poster at NeurIPS 2025.

#Key Features

  • Autonomous multi-scale navigation: The model plans a viewing path over a slide thumbnail, selects regions in a prioritized order, and zooms between 1x, 2x, and 4x crops, integrating observations across views into one diagnosis.
  • Single model across three scales: One multi-stage training run unifies patch-level understanding, large-region reasoning, and whole-slide navigation, rather than stitching together separately trained specialists.
  • Interpretable diagnostic trace: Because each observation is tied to a coordinate and magnification, the generated summary carries an explicit record of what was examined and why, which a pathologist can review.
  • PathMMU-HR2 benchmark: 1,688 expert-validated visual question-answering pairs over 16000x16000-pixel regions spanning ten cancer types, built to require genuine multi-scale analysis.
  • Prompting strategy that transfers: The same navigation protocol applied to general-purpose models lifts their region-level accuracy by an average of 3.2 points, indicating the gain comes from the workflow and not only the weights.

#Technical Details

CPathAgent is built on the LLaVA-OneVision multimodal framework, pairing Qwen3-14B as the language backbone with CPath-CLIP, a pathology-specific vision encoder trained on over 700,000 image-caption pairs, connected by a two-layer MLP projector. Training proceeds in three stages: alignment of the projector alone on patch captions, full fine-tuning on patch-level pathology instructions, and a final stage on the purpose-built CPathAgent-Instruct dataset — region-selection, viewing-path, and multi-image reasoning data distilled from TCGA slide-report pairs with Gemini-2.5-Pro as the annotator — mixed with 20% of the earlier instruction data to preserve patch skill.

On PathMMU the model reaches 80.5% on the tiny test split and 78.6% overall, ahead of Gemini-2.5-Pro at 68.7% / 67.5% and the strongest pathology-specific baseline, CPath-Omni, at 72.4% / 72.2%. On PathMMU-HR2 it scores 88.6% overall against 76.4% for agent-prompted Gemini-2.5-Pro, despite a smaller 1008x1008 single-view resolution. Across six TCGA whole-slide subtyping tasks it averages 82.8% accuracy, versus 79.9% for ABMIL and 72.1% for Gemini-2.5-Pro.

#Applications

The intended setting is diagnostic support in anatomic pathology, where a reviewable navigation trace matters as much as the label. CPathAgent generates region-level descriptions and whole-slide summaries for subtyping tasks across breast, lung, renal, esophageal, bladder, and thyroid cancer, and answers visual questions about specific tissue regions. Pathologist review of 40 regions confirmed that its navigation paths and multi-scale reasoning align with clinical logic, and the pass@k rate rises with repeated generations.

#Impact

CPathAgent shows that the diagnostic workflow itself, not just the encoder, is a modelable object — and that teaching a model to look where a pathologist would look recovers accuracy that raw resolution alone does not buy. The margin over multiple-instance-learning baselines on whole-slide classification is modest at 2.9 points, and the authors report a roughly 9-point gap to the upper bound achievable from their own training data, leaving clear headroom in navigation strategy. Neither code nor model weights have been publicly released, so independent reproduction is not currently possible, and the model is a research prototype without regulatory clearance for clinical use.

At a glance

Released
May 2025
Category
Pathology
Organizations
Westlake University / Zhejiang University / The Ohio State University

Links

Research Paper

Tags

cancer_detectionfoundation_modelhistologymultimodalreport_generationtransformervision_transformervisual_question_answering

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.