BESR links mammography token selection to comparative reasoning across both breasts
Researchers propose Bilateral Evidence Selection and Reasoning, a method that keeps diagnostically useful regions in compressed medical images and rewards models for actually using the opposite-breast comparison.
In plain words
Medical AI models that read mammograms must shrink huge images into a small number of visual tokens, and compressing them can erase the details a radiologist would compare between the two breasts. The BESR method picks which image regions to keep and adds a training reward so the model genuinely reasons with the preserved comparison. The authors report better diagnosis and faster inference under tight token budgets.
What was announced
Medical vision-language models face a compression dilemma: high-resolution scans contain mostly normal anatomy, so trimming visual tokens saves compute, but keeping only salient lesion regions can strip away the context needed to interpret them. The arXiv paper's authors point out that in bilateral mammography this context includes the opposite breast, which serves as a patient-specific anatomical reference. They argue that preserving such evidence alone does not guarantee the model uses it when judging findings.
Their Bilateral Evidence Selection and Reasoning approach pairs a bilateral-guided token selector with a bilateral reward. Under a shared token budget, the selector weighs lesion appearance together with its local and cross-breast context, while the reward pushes the model toward assessments that benefit from the preserved comparison. On a bilateral dataset built from CBIS-DDSM and on VinDr-Mammo, the team reports improved evidence retention and diagnostic performance under constrained token budgets, plus reduced prefill latency in inference benchmarking.
What it does not show
- Evidence is limited to the paper's abstract; the full article was not reviewed, so experimental details and comparisons cannot be verified.
- The reported gains come from the authors' own experiments on two datasets and have not been independently confirmed.
For security and IT teams
- Scope
- The method targets bilateral mammography analysis under constrained visual-token budgets, not general medical imaging.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.