Value of deep learning model for predicting Breast Imaging Reporting and Data System 3 and 4A lesions on mammography
Xiaohui Lin1, Tingting Liao1, Yuting Yang1
1Department of Radiology, Shenzhen People's Hospital, The Second Clinical Medical College, Jinan University, Shenzhen, China.
Background:
The diagnostic categorization of a lesion as Breast Imaging Reporting and Data System (BI-RADS) category 3 or 4A determines whether a patient needs a biopsy; however, interobserver variability exists among radiologists in mammographic interpretation. This variability may lead to underdiagnoses of BI-RADS 3 lesions and unnecessary biopsies of benign BI-RADS 4A lesions. Therefore, we assessed the diagnostic value of a mammography-based deep learning (DL) model for differentiating BI-RADS 3 and 4A lesions and its impact on radiologists' decision-making.
Methods:
This retrospective multicenter study analyzed 846 mammographically detected breast lesions (BI-RADS 3 and 4A) from 824 patients at Shenzhen People's Hospital and Shenzhen Luohu People's Hospital between January and December 2020. Six breast imaging specialists (three junior and three senior) independently reviewed all mammograms with and without DL model assistance. The follow-up or biopsy results were used as the reference standard. Diagnostic performance was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC), with the DeLong test being used to compare AUCs.
Results:
The DL model yielded an AUC of 0.74 for distinguishing BI-RADS 3 and 4A lesions, outperforming junior radiologists' standalone performance (AUC =0.57, AUC =0.55, and AUC =0.58) but remaining inferior to senior radiologists (AUC =0.78, AUC =0.77, and AUC =0.76). With DL model assistance, all six radiologists had higher AUCs for diagnosing BI-RADS 3 and 4A lesions as compared to their unassisted performance. Importantly, DL integration significantly increased junior radiologists' AUCs to 0.74-0.77 (P<0.001), whereas the increase in the AUCs of the senior radiologists (to 0.79, 0.78, and 0.78) was not significant (P>0.05).
Conclusions:
The mammography-based DL model significantly improved the diagnostic performance of junior radiologists for BI-RADS 3 and 4A lesions, effectively reducing missed diagnoses and unnecessary biopsies.
