How Do We Evaluate AI That Locates and Explains Findings on MRI?

The 7,970 MRI volumes were split into 5,977 training, 797 validation and 1,196 test volumes without patient overlap.

The dataset links MRI volumes, diagnostic questions, 3D locations, clinical reasoning and severity. Spatial localisation and diagnostic reasoning are evaluated separately.
Pacific Symposium on Biocomputing (2026). DOI: 10.1142/9789819824755_0008 · Original explanatory chart based on reported data
Evidence & Original Source Image use terms

Research structure

01
MRI footage.
3-dimensional volume
02
Diagnostic questions
Sector of interest and disease
03
3D location
Bounding box
04
Doctor's explanation
The stage of judgment.
05
It's a pain in the ass.
Structured evaluation

MRI images, diagnostic questions, 3D locations, stages of the physician’s explanation, and severity assessments were combined into one dataset. Localization and diagnostic performance were evaluated separately.

Explanatory schematic

Evidence & Original Source

The research question

Can a 3D MRI Dataset Evaluate Both Anatomical Localization and Step-by-step Diagnostic Reasoning?

Participants and Methods

The dataset combined MRI volumes, diagnostic questions, 3D bounding boxes, physicians’ reasoning steps and severity assessments. Five VLMs were compared on ReasonKnee-Bench. Approximately 494,000 data sets were constructed from 7,970 knee MRI volumes. Expert annotation required more than 450 hours.

Main Findings

The 7,970 MRI volumes were split into 5,977 training, 797 validation and 1,196 test volumes without patient overlap. Approximately 494,000 tuples link expert reasoning, location and severity. Mean Dice for anatomical subregion segmentation was 0.717.

Implications for clinical and surgical practice

The study provides research data for AI that interprets MRI. It does not establish the effectiveness of autonomous intraoperative decision-making.

Limitations

This is a dataset and benchmark study. It did not directly evaluate accuracy in clinical practice or treatment and surgical outcomes. The 2025 arXiv preprint and the 2026 conference proceedings are different publication versions of the same study. The supplied manuscript is the 2025 preprint. Methods and numerical results may differ in the final conference proceedings.

Let's put it briefly.

Which clinical question does this address?

The study provides research data for AI that interprets MRI. It does not establish the effectiveness of autonomous intraoperative decision-making.

Keywords

Imaging, Three-Dimensional
Magnetic Resonance Imaging
Computational Biology
Knee
Artificial Intelligence
AI
MRI
vision-language models

Abstract · Original text

Sraavya Sambara, Sung Eun Kim, Xiaoman Zhang, Luyang Luo, Shreya Johri, Mohammed Baharoon, Du Hyun Ro, Pranav Rajpurkar. 3DReasonKnee: Advancing Grounded Reasoning in Medical Vision Language Models. Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing. 2026;31:99-113. doi:10.1142/9789819824755_0008.

Excerpt from the original abstract

Current Vision-Language Models (VLMs) struggle to ground anatomical regions in 3D medical images and reason about them in a step-by-step manner-a key requirement of realworld…

Read abstract on PubMed

The original paper.

3DReasonKnee: Advancing Grounded Reasoning in Medical Vision Language Models

Sraavya Sambara, Sung Eun Kim, Xiaoman Zhang, Luyang Luo, Shreya Johri, Mohammed Baharoon, Du Hyun Ro, Pranav Rajpurkar
Pacific Symposium on Biocomputing · 2026

DOI · 10.1142/9789819824755_0008 PubMed Based data

Reviewed academic papers

Research question and design