SPARC-Rad: New Benchmark for Spatial and Anatomical Reasoning in Radiology VLMs
A team of researchers has launched a new dataset and evaluation framework dubbed SPARC-Rad, aimed at enhancing the assessment of spatial and anatomical reasoning in radiology vision-language models. This innovative dataset features 300 image-question pairs sourced from The Cancer Imaging Archive (TCIA), including various modalities such as CT, MRI, and radiography, covering areas like the abdomen, chest, breast, neuro, and musculoskeletal systems. Questions designed by radiology trainees focus on anatomical recognition and spatial reasoning. The initiative seeks to boost the practical application of vision-language models in clinical settings and is accessible on arXiv, entry 2608.00100.
Key facts
- SPARC-Rad is a benchmark for spatial and anatomical reasoning in radiology VLMs.
- It includes 300 image-question pairs from healthy control imaging studies in TCIA.
- Imaging modalities: CT, MRI, and radiography.
- Anatomical categories: abdomen, chest, breast, neuro, and musculoskeletal.
- Questions were manually designed and annotated by radiology trainees.
- Evaluates anatomical identification, localization, laterality, regional recognition, device identification, and inter-structure spatial relationships.
- Available on arXiv with ID 2608.00100.
Entities
Institutions
- The Cancer Imaging Archive (TCIA)
- arXiv