BreastViT + GPT-4o generates structured breast ultrasound reports with top radiologist scores

Combined BreastViT+GPT-4o achieved top radiologist-rated scores for structural integrity (8.03), terminology (8.02), accuracy (7.31) and completeness (7.61) on breast US, significantly outperforming standalone models (P<0.05) in external validation (49 cases).

• Retrospective study: 2,243 images (362 patients) for training; external validation on 49 BI-RADS 4B-5 cases.
• Three senior radiologists' blinded evaluation showed excellent inter-rater agreement (ICC 0.88, 95% CI 0.86–0.90).
• Single-center, retrospective design; limited external dataset; real-world workflow integration not yet assessed.

RadPigeon summaries are original and for information only. They are not clinical advice.