MBZUAI benchmark tests spatial reasoning in 3D medical AI
MBZUAI’s CT-SpatialVQA benchmark tests how medical AI models identify spatial relationships in CT scans.
Event / data period:
Published: · Updated:
Researchers at MBZUAI developed CT-SpatialVQA, a benchmark for testing whether 3D medical AI systems can track the spatial relationships of anatomy and abnormalities across computed tomography (CT) volumes. In a 19 October university report, the dataset comprised 9,077 spatial question-and-answer pairs derived from 1,601 chest CT scans and radiology reports.
The researchers evaluated eight vision-language models in a zero-shot setting. None reached 50% accuracy, and the average across the eight was 34.08%. CT-Chat scored 43.69%, MedEvalKit 40.16%, MedGemma-1.5 38.24% and VILA-M3 28.20%. The work was accepted as an oral presentation at MICCAI 2026. These are results from a particular research benchmark; they do not establish clinical competence or validated use in patient care.
Data
CT-SpatialVQA model accuracy results
Selected results and overall mean from the university’s evaluation of eight models.
| CT-Chat | 43.69 | percent |
| MedEvalKit | 40.16 | percent |
| MedGemma-1.5 | 38.24 | percent |
| VILA-M3 | 28.2 | percent |
| Average across eight models | 34.08 | percent |
Methodology
These are zero-shot accuracy results on the CT-SpatialVQA benchmark. The mean covers eight models; the table lists four models reported individually in the source and the mean. Results are not clinical validation.
Updated:
CT-SpatialVQA dataset
Question and scan coverage in the research benchmark.
| Spatial question-answer pairs | 9077 | count |
| Chest CT scans | 1601 | count |
Methodology
Counts describe the researchers’ test benchmark, not a patient population, diagnostic performance or clinical outcome. The source says the 9,077 questions were derived from 1,601 chest CT scans and radiology reports in the test set.
Updated:
Subjects and places
Subject proposals
- Artificial intelligenceSuggested
- Subject suggested from the title or summary: AImbzuai.ac.ae ↗
- HealthSuggested
- Subject suggested from the title or summary: medicalmbzuai.ac.ae ↗
Geographic references
Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.
- United Arab Emirates · Place role: Coverage · Place type: Country or area · Location precision: RegionalEvidence
- Explicit preserved editorial country/region; no city or national centroid inferredmbzuai.ac.ae ↗