A new AI system for answering questions about three-dimensional breast MRI reported a small accuracy edge and lower measured latency than a baseline given the full visual input. At a prototype budget of 512, SeVeR reported 70.21% accuracy in 644 milliseconds, compared with 69.18% accuracy and 953 milliseconds for the full-token baseline.
The result comes from an arXiv preprint dated 26 August 2026. The study evaluates whether selecting and retrieving some of the visual information in a multi-sequence scan can reduce redundant 3D visual-token exposure while preserving or improving question-answering performance. Its evidence is algorithmic benchmark testing, not an assessment of patient outcomes or radiology workflow performance.
A large benchmark built from breast MRI records
The BreMRIs-VQA benchmark contains 71,041 breast MRI sequences from 12,891 patients. Each patient contributed at least three sequences across six MRI modalities, and the scans were paired with expert-verified radiology and pathology reports.
The released benchmark contains 1,186,726 volumetric question-and-answer pairs. Of these, 671,596 are free-text pairs and 515,130 are multiple-choice pairs spanning seven workflow-grounded tasks. Cases were grouped into strata, shuffled with fixed random seed 42, split 96% for training and 4% for testing, and kept disjoint between training and evaluation.
The system chooses what to expose
SeVeR is built around selective visual exposure and retrieval. Its Greedy Prototype Selection compresses each MRI modality; Change-aware Gated Attention performs question-conditioned, multi-level retrieval; and Self-Consistency Regularization with Marginal Utility penalizes retrieval that fails to improve task loss.
On BreMRIs-VQA, the SeVeR-4B configuration paired with Qwen3-VL-4B recorded an average accuracy of 70.57 and an average BERTScore of 98.70 across seven tasks.
Strong scores on other tests
The paper also reports results on the public 3D-RAD benchmark. SeVeR recorded accuracies of 82.62 for Existence Detection, 51.77 for Static Temporal Diagnosis and 75.28 for Longitudinal Temporal Diagnosis. The paper reports these as the highest values for the three tasks.
On DeepTumorVQA, SeVeR-4B had the highest overall averages reported for multiple-choice and free-text evaluation, at 0.687 and 0.599. The reported margins over the second-best system were 0.048 and 0.060 in Visual Reasoning, and 0.072 and 0.034 in Medical Reasoning, for the two evaluation formats respectively.
The fixed-budget comparison also lists lower scores for versions without individual components. At 512 tokens, the version labeled w/o CaGA scored 64.36 in accuracy versus 70.57 for SeVeR, while the version labeled w/o Regul. scored 87.04 in BLEU versus 91.08 for SeVeR.
Reported average accuracy values were 70.21% with Qwen2.5-VL-3B, 70.57% with Qwen3-VL-4B and 72.13% with Qwen3-VL-8B. In tests with missing or substituted modalities, SeVeR showed smaller performance gaps relative to the all-sequence S6 setting than the w/o Proto baseline across the tested settings and clinical tasks.
What remains untested
The benchmark questions underwent quality control using Qwen3-Coder-Plus and a radiologist audit of 3,500 question-and-answer pairs, with 500 pairs sampled per task. The retained set had an average overall score of 2.85 out of 3, and 97.5% of audited pairs passed expert checking.
The study's free-text results rely on automated language metrics, which may not capture clinical nuance. Radiologist-grounded human evaluation was left for future work. The paper also did not evaluate adaptive prototype budgets based on sequence complexity or integration into real-world radiology workflows.
The findings are limited to algorithmic comparisons on the report-derived BreMRIs-VQA benchmark and the public 3D-RAD and DeepTumorVQA benchmarks. They do not establish improved clinician decisions, patient outcomes, diagnostic safety or benefit in real-world radiology workflows.
The retrospective benchmark construction was institutionally approved and the clinical images and reports were de-identified. Source images, reports and other restricted artifacts will not be openly distributed.
Funding
The work reports partial support from the National Natural Science Foundation of China, Zhejiang R&D programs, Guangzhou City's Key R&D Program, DAMO Academy and a Zhejiang postdoctoral program.
Paper data and sources
Original title: SeVeR: Selective Visual Exposure and Retrieval for 3D Medical Image Question Answering
Authors: Yaojun Hu, Danyang Tu, Yang Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text