CaseScope: A Multi-Field Benchmark for Case-Level Pathology Diagnosis with Multimodal LLMs
Paper • 2611.15325 • Published
Datasets:
Pathologists rarely diagnose from a single image: they scan a section at low power for architecture, zoom into selected fields for cytology, and integrate the evidence across views and scales. CaseScope evaluates multimodal large language models (MLLMs) on this setting: each item presents several curated H&E fields of one lesion (for atlas figure series, of one disease), ordered from low to high magnification, and asks the model to diagnose the case, triage its behaviour, grade it, verify findings or locate one.
✨ Highlights
- 🔬 Case-level, multi-field: 2 to 8 fields per case, ordered from low to high magnification, instead of a single image or a whole-slide pipeline
- 🏥 Two complementary sources: a residency-training image library and 88 published pathology atlases
- 🧬 Broad diagnostic coverage: 843 entities across 18 organ systems, including non-neoplastic, benign and rare entities
- 🧩 Six tasks, 4,679 items: open-ended and ten-way diagnosis, behaviour triage, ordinal grading, finding verification and finding location
- 🎯 Hard and robust: ontology-guided distractors that include each entity's classic look-alikes, CircularEval over option rotations, field-set ablations
- 🌐 Bilingual: questions, options and instructions in English and Chinese
📊 At a Glance
| 🏥 CaseScope-Teaching | 📚 CaseScope-Atlas | |
|---|---|---|
| Source | in-house residency-training image library | 88 published pathology atlases |
| Cases | 436 | 800 (one per entity) |
| Fields shown to the models | 2,277 | 2,664 |
| Fields per case | median 5 (2 to 8) | median 3 (2 to 8) |
| Entities | 390 | 800 (453 without a Teaching case) |
| Behaviour (non-neoplastic / benign / intermediate / malignant) | 114 / 96 / 61 / 165 | 210 / 190 / 107 / 293 |
| H&E-sufficiency prior (Y / P / N) | 194 / 221 / 21 | 176 / 601 / 23 |
| Specimen site stated in the question | 400 | 748 |
| Items | 1,686 | 2,993 |
🫀 Cases per organ system
| Organ system | Teaching | Atlas |
|---|---|---|
| Soft tissue | 28 | 121 |
| Female genital | 56 | 78 |
| Skin | 21 | 91 |
| GI tract | 44 | 52 |
| Lymph node & spleen | 23 | 56 |
| Breast | 35 | 42 |
| Liver, biliary & pancreas | 31 | 45 |
| Urinary | 19 | 44 |
| Lung | 25 | 37 |
| Oral & salivary | 30 | 31 |
| Nervous system | 12 | 49 |
| Male genital | 22 | 28 |
| Mediastinum & pleura | 25 | 23 |
| Endocrine | 18 | 28 |
| Bone & joint | 13 | 29 |
| Sinonasal & larynx | 18 | 21 |
| Ear & eye | 8 | 18 |
| Peritoneum | 8 | 7 |
🧪 Tasks
| Task | What the model does | Teaching | Atlas | Answer | Metric |
|---|---|---|---|---|---|
| A1 | Open-ended diagnosis | 436 | 800 | JSON: diagnosis, two differentials, confidence | hierarchical diagnostic score (HDS), exact, top-3 |
| A2 | Diagnosis, 10-way multiple choice | 436 | 800 | option letter | CircularEval, 10 rotations |
| B | Behaviour triage, 4-way | 380 | 762 | option letter | CircularEval, critical miss rate |
| C | Ordinal grading (e.g. DCIS nuclear grade, Gleason pattern, endometrial dating) | 44 | - | option letter | CircularEval, quadratic weighted kappa |
| R1 | Verify four statements on the low-power architecture or the high-power cytology | 390 | 195 | four booleans | statement and item accuracy |
| G1 | Locate a described finding among the fields | - | 436 | option letter (a field) | CircularEval over the option rotations |
CircularEval: a multiple-choice item is stored once per rotation of its options (rows sharing a
base_id) and counts as correct only if every rotation is answered correctly. HDS: 1 for the reference entity, a synonym or an accepted alternate; 0.5 for another entity of the same family and behaviour; 0 otherwise.
🚀 Quick Start
Download
from huggingface_hub import snapshot_download
root = snapshot_download("WenchuanZhang/CaseScope", repo_type="dataset")
Read an item and its fields
import json, os
from PIL import Image
item = json.loads(open(os.path.join(root, "teaching", "A2.jsonl")).readline())
fields = [Image.open(os.path.join(root, "images", "teaching", f"{i}.jpg")) for i in item["images"]]
print(item["question_en"])
for o in item["options"]:
print(f"{o['key']}. {o['en']}")
The text fields of each task can also be browsed and loaded as a configuration, e.g.
load_dataset("WenchuanZhang/CaseScope", "teaching_A2", split="train"); the fields are referenced by id.
Evaluate a model with the evaluation code: any multi-image model supported by vLLM or any OpenAI-compatible endpoint.
python -m casescope.infer --data <dataset folder> --model Qwen/Qwen2.5-VL-7B-Instruct
python -m casescope.judge --pred results/Qwen2.5-VL-7B-Instruct --data <dataset folder>
python -m casescope.score --pred results/Qwen2.5-VL-7B-Instruct --data <dataset folder>
📁 Dataset Structure
CaseScope/
├── images/
│ ├── teaching/I*.jpg 2,277 fields (JPEG, long side ≤ 1,024 px)
│ └── atlas/IA*.jpg 2,664 fields
├── teaching/ A1, A2, B, C, R1 (.jsonl)
├── atlas/ A1, A2, B, R1, G1 (.jsonl)
└── metadata/
├── cases_teaching.jsonl 436 cases
├── cases_atlas.jsonl 800 cases
└── ontology.csv 846 entities
Example item (teaching/A2.jsonl, shortened)
{
"task": "A2",
"item_id": "A2-C0001-r0",
"base_id": "A2-C0001",
"case_id": "C0001",
"images": ["I0347", "I0879", "I0744", "..."],
"question_en": "The following 7 H&E-stained microscopy images are fields from the same lesion of one case, ordered from low to high magnification. Specimen site: skin. What is the most likely pathological diagnosis?",
"options": [
{"key": "A", "en": "Solitary fibrous tumour"},
{"key": "B", "en": "Junctional melanocytic naevus"},
{"key": "C", "en": "Melanoma"},
"..."
],
"instruction_en": "Integrate all images and answer with the option's letter from the given choices directly.",
"answer": "E",
"rotation": 0,
"meta": {"entity_id": "blue_nevus", "behavior": "B", "he_sufficiency": "Y", "n_images": 7}
}
📋 Item fields
| Field | Description |
|---|---|
task, item_id, base_id, case_id |
identifiers; the rotations of one item share its base_id |
images |
field ids in presentation order (low to high magnification) |
question_en, question_zh |
the question, with the number of fields and the specimen site |
options |
key, en, zh for each option; null for A1 and R1 |
instruction_en, instruction_zh |
the answer instruction, placed after the options |
answer |
option key; entity id for A1; four booleans for R1 |
rotation |
index of the option rotation (multiple-choice tasks) |
meta.entity_id, meta.behavior |
reference entity; NN non-neoplastic, B benign, I intermediate, M malignant |
meta.he_sufficiency |
Y diagnosable from H&E alone, P suggestive, N ancillary tests required |
meta.n_images |
number of fields |
meta.distractors, meta.lookalikes |
A2: the distractor entities and the classic look-alikes among them |
meta.group, meta.level, meta.n_levels |
C: ordinal group and the reference level |
meta.level |
R1: low (architecture) or high (cytology) |
meta.tier, meta.same_case |
Atlas: T1 common diagnosis, T2 other WHO entity, T3 other diagnosis; whether the fields show one lesion (true, 413 cases) or one disease (false, 387 cases) |
📋 Case and ontology fields
| File | Fields |
|---|---|
cases_*.jsonl |
case_id, entity_id, name_en, name_zh, behavior, he_sufficiency, organ_system, site_en, site_zh (empty when no site is stated), alternates (accepted alternate diagnoses), image_ids; Atlas also tier, same_case |
ontology.csv |
entity_id, name_en, name_zh, synonyms (separated by |), behavior, category, lineage, family, he_sufficiency, added_for_atlas (CaseScope-Teaching is scored against the original 407 entities, CaseScope-Atlas against all 846) |
📈 Results
Main results of the paper (21 MLLMs)
Scores in %. A2, B, C, G1: CircularEval; A1: HDS; R1: statement accuracy; the best open and the best proprietary model per column in bold. †Single-image model given a numbered mosaic of the fields.
| Model | Teaching A1 | A2 | B | C | R1 | Atlas A1 | A2 | B | R1 | G1 |
|---|---|---|---|---|---|---|---|---|---|---|
| Proprietary MLLMs | ||||||||||
| Claude Sonnet 5.5 | 64.2 | 61.7 | 75.5 | 79.5 | 93.2 | 41.6 | 49.1 | 61.5 | 80.8 | 61.5 |
| Claude Haiku 5.5 | 43.9 | 39.0 | 57.4 | 54.5 | 83.2 | 28.8 | 29.9 | 47.0 | 73.6 | 48.2 |
| GPT-5.6-sol | 43.5 | 49.3 | 55.5 | 61.4 | 92.9 | 27.6 | 37.1 | 41.9 | 78.6 | 45.6 |
| GPT-5.6-luna | 33.5 | 31.9 | 51.1 | 56.8 | 87.1 | 19.8 | 24.1 | 32.7 | 73.1 | 39.9 |
| General MLLMs (open) | ||||||||||
| Qwen2.5-VL-7B | 11.7 | 6.0 | 18.7 | 9.1 | 63.0 | 10.9 | 9.8 | 17.2 | 63.6 | 39.4 |
| InternVL3.5-8B | 18.3 | 11.0 | 30.5 | 27.3 | 66.5 | 12.3 | 9.4 | 19.9 | 67.6 | 36.7 |
| Qwen3-VL-8B | 11.9 | 10.6 | 30.5 | 27.3 | 63.3 | 12.3 | 12.1 | 17.1 | 64.6 | 32.8 |
| Qwen3.5-9B | 29.7 | 17.9 | 31.8 | 36.4 | 81.5 | 18.6 | 14.2 | 28.1 | 69.5 | 42.7 |
| MiMo-VL-7B | 15.9 | 7.8 | 26.8 | 29.5 | 66.1 | 12.2 | 8.9 | 23.6 | 67.3 | 33.5 |
| GLM-4.1V-9B-Thinking | 19.4 | 12.6 | 31.1 | 29.5 | 74.7 | 15.4 | 12.6 | 31.6 | 68.3 | 36.2 |
| Medical MLLMs | ||||||||||
| HuatuoGPT-Vision-7B (Q2.5-VL) | 12.0 | 8.0 | 30.8 | 13.6 | 57.8 | 11.8 | 9.4 | 24.7 | 64.9 | 31.7 |
| Lingshu-7B | 21.0 | 5.7 | 34.2 | 36.4 | 57.8 | 14.8 | 7.0 | 32.3 | 61.3 | 39.4 |
| OctoMed-7B | 14.1 | 4.8 | 12.9 | 20.5 | 62.2 | 8.4 | 4.8 | 15.1 | 60.4 | 22.5 |
| MedGemma-1.5-4B | 11.5 | 2.5 | 18.9 | 20.5 | 51.7 | 5.9 | 1.8 | 17.7 | 57.8 | 11.0 |
| Hulu-Med-7B | 17.3 | 13.3 | 33.4 | 13.6 | 63.3 | 9.6 | 4.8 | 20.7 | 63.2 | 21.6 |
| Pathology MLLMs | ||||||||||
| Quilt-LLaVA-7B† | 3.8 | 1.6 | 1.1 | 2.3 | 49.4 | 2.0 | 0.6 | 10.6 | 46.4 | 15.1 |
| PathGen-LLaVA-13B† | 17.4 | 4.4 | 13.9 | 4.5 | 50.1 | 8.5 | 1.9 | 22.8 | 49.7 | 16.1 |
| CLOVER | 12.5 | 4.1 | 20.5 | 36.4 | 59.9 | 11.6 | 7.2 | 14.2 | 63.6 | 16.1 |
| Patho-R1-3B | 13.4 | 0.2 | 2.4 | 6.8 | 12.1 | 14.3 | 0.1 | 3.1 | 14.4 | 2.1 |
| Patho-R1-7B | 34.7 | 7.1 | 33.2 | 29.5 | 36.1 | 21.5 | 7.6 | 19.4 | 32.2 | 20.6 |
| ScaleReasoner-R1 | 32.5 | 10.8 | 34.5 | 25.0 | 29.2 | 20.1 | 8.6 | 28.7 | 26.3 | 18.8 |
🛠️ Construction
🏥 CaseScope-Teaching. The library was built by selecting common diseases from pathology textbooks, scanning the corresponding slides and having four pathologists collect fields at several magnifications and jointly select the representative ones. Copies of a picture filed under several organ systems were found by file and perceptual hashes and pathology-CLIP similarity confirmed at the pixel level, and merged into cases. All images were de-identified (no EXIF, ICC or XMP metadata; opaque identifiers). Labels were harmonised to an ontology aligned with the WHO Classification of Tumours (5th edition); an automated pre-review flagged doubtful cases and items for re-examination by a pathologist, and a pathologist checked the field order of every case. The R1 statements were generated by an LLM from the label and the WHO criteria and reviewed by a pathologist.
📚 CaseScope-Atlas. H&E panels sharing a caption form a case of one lesion, and consecutive figures of one entity a case of one disease. Each case was mapped from its caption to the ontology, extended by 439 entities, by two LLMs that had to agree, and the labels were verified by a pathologist; one case per entity was kept, common diagnoses first. The R1 statements and G1 findings are derived from the captions; the captions themselves are not distributed.
📜 License
CC BY-NC 4.0: free to share and adapt for non-commercial purposes with attribution. The evaluation code is MIT-licensed.
📝 Citation
@article{zhang2026casescope,
title = {{CaseScope}: A Multi-Field Benchmark for Case-Level Pathology Diagnosis with Multimodal {LLMs}},
author = {Zhang, Wenchuan and Zhang, Chengyang and Li, Fengling and Yin, Ting and Shu, Chen and Chen, Jie and
Zhang, Shuwan and Wu, Xiaoyan and Zhang, Zhang and Yi, Yuhao and Lv, Jiancheng and Bu, Hong},
journal = {arXiv preprint arXiv:2611.15325},
year = {2026}
}
- Downloads last month
- 125