Deep learning–based computer-aided diagnosis on breast ultrasound may significantly improve radiologists’ diagnostic performance and reduce the frequency of benign breast biopsies, according to a new study published by He et al in the American Journal of Roentgenology. Compared with previous studies supporting computer-aided diagnosis at tertiary and/or urban centers, the new findings involving radiologists without breast ultrasound expertise “support [computer-aided diagnosis] in settings with incomplete access to breast imaging expertise,” explained senior study author Li-Gang Cui, MD, Professor and Chair of the Department of Ultrasound at the Peking University Third Hospital.
Study Methods and Results
In the new prospective multicenter study, the researchers included patients scheduled to undergo biopsies or surgical resections of breast lesions that were classified as Breast Imaging-Reporting and Data System (BI-RADS) categories 3 to 5 on prior breast ultrasounds from November 2021 to September 2022. The patients were assigned to undergo additional investigational breast ultrasounds, both performed and interpreted by radiologists with no expertise in the modality.
Hybrid body-breast imagers—radiologists lacking breast subspecialty training or those whose breast ultrasounds accounted for less than 10% of the ultrasounds they performed annually—then assigned a BI-RADS category to each breast lesion. Computer-aided diagnosis results were used to upgrade reader-assigned BI-RADS category 3 lesions to category 4A as well as to downgrade BI-RADS category 4A lesions to category 3. Histologic results of the biopsies or surgical resections served as the researchers’ reference standard.
Ultimately, the application of computer-aided diagnosis to the interpretations made by radiologists without breast ultrasound expertise resulted in an upgrade of 6.0% (n = 6/100) of BI-RADS category 3 assessments to category 4A, 16.7% (n = 1/6) of which were found to be malignant. Further, the use of computer-aided diagnosis led to the downgrade of 79.1% (n = 87/110) of category 4A assessments to category 3, 4.6% (n = 4/87) of which were malignant.
Conclusions
Because institutions lacking breast imaging expertise may also suffer from capacity issues to perform image-guided breast biopsies and pathologic evaluations of biopsy specimens, computer-aided diagnosis could be used to decrease the rate of benign breast biopsies.
Disclosure: For full disclosures of the study authors, visit ajronline.org.
