Wavelet-Enhanced Deep Learning Cuts Unnecessary Biopsies in BI-RADS 4a+ Breast Lesions

A retrospective study (n=390) found a wavelet-spectrum multi-channel deep learning model could spare 26% of biopsies in BI-RADS 4a+ breast lesions while keeping missed malignancy at 1.8% — a ~2% accuracy gain over B-mode alone. Not yet externally validated.

• Retrospective single-center study of 390 patients; 3-channel ultrasound (B-mode + 13 MHz + 7 MHz wavelet components) fed into a dual-channel deep learning classifier; overall accuracy 84.52%.
• Patient-specific biopsy-sparing framework yielded a 26.01% biopsy-sparing rate with 1.786% missed diagnosis rate at the most stringent threshold; biopsy-sparing-to-missed-diagnosis ratio remained robust across thresholds.
• Key limitation: retrospective design, small cohort (n=390), no external or prospective validation — limits generalizability before clinical adoption.

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