A breast-imaging preprint examines whether combining mammograms and radiology reports can help identify benign tumours among patients recommended for biopsy. The authors focus on BI-RADS 4 and 5 cases and add visual explanations to the model’s predictions.
Two kinds of evidence
The study uses an in-house dataset of 1,632 patients, each with four mammography views and a corresponding report. A Vision Transformer extracts image features, while DeBERTa-V3-Large produces text embeddings. The system combines the independently extracted features through late fusion before classification.
Grad-CAM highlights image regions that influence a prediction. This provides something a reviewer can inspect, although a highlighted region by itself does not establish that the model’s reasoning is clinically sound or that its recommendation is safe.
Reported discrimination and thresholds
The best reported test-set AUC is 0.72 for the multimodal model. Image-only and text-only baselines achieved 0.66 and 0.68 respectively. These figures describe discrimination on the study’s data, rather than the clinical effect of changing biopsy decisions.
The authors also report sensitivity settings of 90%, 95%, 99% and 100%, with corresponding specificities of 40%, 36%, 18% and 11%. Preserving those pairs matters: an operating point that catches more positive cases can leave fewer benign cases correctly identified as negative.
A possible research direction for biopsy review
The paper proposes that multimodal information and explanations could support efforts to reduce unnecessary biopsies. It does not report a prospective programme in which biopsies were safely avoided, nor establish improved patient outcomes.
The next useful evidence would examine performance in additional patient populations and evaluate a clearly specified decision threshold. Any clinical comparison would need to account for missed disease and the consequences of a changed biopsy recommendation, alongside the number of procedures potentially avoided.
Comments