Oncology Report-Derived VQA - Brain MRI RADS-Style: leaderboard
Zero-shot multiple-choice visual question answering on brain MRI of confirmed cancer patients (UCSF-PDGM images paired with private institutional reports), with questions derived deterministically from a clinician-validated RANO-aligned reporting schema (Liu et al., 2026). The model sees one representative slice with the question and options, never the report. Accuracy in percent over valid responses; higher is better.
Metric: Accuracy (%; zero-shot multiple-choice VQA with 2-5 options on one representative 2D slice of a 3D study, the model sees the image, question and options but never the report; invalid responses excluded; brain MRI, 480 cases (UCSF-PDGM images with private reports), 4,174 RADS-style questions from a RANO-aligned schema). Source: arxiv.org. Saturation forecast: Around January 2028. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 69.3 | #100 |
| 2 | GPT-5.2 | 65.8 | #172 |
| 3 | GPT-5 | 60.7 | #163 |
| 4 | MedGemma 1.5 4B | 55.6 | #2579 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=oncology-report-derived-vqa-brain-mri-rads-style · How It Works · Data refreshed daily, snapshot 2026-11-12.