Scale-VQA: Cross-scale Pathology Visual Question Answering

MICCAI 2026 | Project Page | Paper | Code | Model

Scale-VQA is a cross-scale pathology VQA dataset for evaluating reasoning across low-, medium-, and high-magnification images. Each question is grounded in multi-magnification pathology images and uses multiple-choice answers.

Dataset contents

This repository contains:

• 3,230 training samples in train.json
• 450 validation samples in val.json
• 1,005 test samples in test.json
• Referenced pathology ROI images organized by WSI identifier

The dataset has 4,685 questions across 15 organs and five reasoning dimensions: correspondence, confirmation, localization, explanation, and diagnosis.

Repository structure

Scale-VQA/
├── train.json
├── val.json
├── test.json
└── <wsi_name>/
    └── rois/
        ├── 10_*.jpg or 10_*.png
        ├── 40_*.jpg or 40_*.png
        └── 200_*.jpg or 200_*.png

Image paths in each JSON file are relative to the dataset repository root. For example, the path below resolves to an image in this repository:

01K46JR67HPRVBYFQ9QK4H2Z12/rois/10_58.53226y_00000x.jpg

Most samples contain low-, medium-, and high-magnification images. Some samples have null for high_mag because a corresponding high-magnification ROI is not available.

Sample format

{
  "question": "...",
  "options": {
    "A": "...",
    "B": "...",
    "C": "...",
    "D": "..."
  },
  "answer": "D",
  "rationale": "...",
  "image_path": {
    "low_mag": "wsi_name/rois/10_....jpg",
    "mid_mag": "wsi_name/rois/40_....jpg",
    "high_mag": "wsi_name/rois/200_....jpg"
  },
  "extra_info": {
    "wsi_name": "...",
    "path_id": "...",
    "question_type": "type_E"
  }
}

Implementation

The preprocessing, data-loading, training, and evaluation code is maintained in the ScaleReasoner-R1 repository. Refer to that repository for the training setup, prompt formatting, RL configuration, and implementation details.

Citation

@misc{phan2026scalereasoner,
  title         = {Enhancing Pathological VLMs with Cross-scale Reasoning},
  author        = {Phan, Chi and Zhang, Tianyi and Xue, Qiaochu and Wu, Yufeng and Hu, Dan and Liu, Zeyu and Wang, Sudong and Jin, Yueming},
  year          = {2026},
  eprint        = {2606.17412},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}