Computer-aided diagnosis aids lesion classification in breast biopsies

The use of Computer-aided diagnosis (CAD) greatly enhanced radiologists' diagnostic abilities, with a potential to decrease the incidence of benign breast biopsies in particular, says an article published in American Journal of Roentgenology.

Radiologists with experience in breast ultrasonography have mostly tested computer-aided diagnostic programs for breast ultrasound interpretation at tertiary and/or metropolitan medical centers. Ping He and team conducted this study to assess the effectiveness of deep learning-based CAD software on the diagnostic performance of radiologists at secondary or rural hospitals without breast ultrasound expertise in differentiating benign from malignant breast lesions measuring up to 2.0 cm on ultrasound.

Patients scheduled to have a biopsy or surgical excision of a breast lesion that had been previously diagnosed as BI-RADS type 3-5 on breast ultrasonography between November 2021 and September 2022 at any of eight participating secondary or rural hospitals in China were included in this prospective research. Patients underwent a second investigational breast ultrasound, which was conducted and interpreted by a hybrid body-breast radiologist (who either lacked breast imaging subspecialty training or in whom the number of breast ultrasounds performed annually accounted for less than 10% of all ultrasounds performed by the radiologist annually) who assigned a BI-RADS category. Reader-assigned BI-RADS category 3 lesions were downgraded to category 4A using CAD findings, while reader-assigned BI-RADS category 4A lesions were upgraded to category 3.

The key findings of this study were:

1. The study comprised 313 individuals with 313 breast lesions (102 malignant and 211 benign), with a mean age of 47.0 14.0 years.

2. 6.0% (6/100) of BI-RADS category 3 lesions were upgraded by CAD to category 4A, with 16.7% (1/6) of those being malignant. 7.9% (87/101) of category 4A lesions were lowered to category 3 by CAD, of which 4.6% (4/87) were malignant.

3. In terms of accuracy (86.6% vs 62.6%; p.001), specificity (82.9% vs 46.0%; p.001), and positive predictive value (PPV; 72.7% vs 46.5%; p.001), diagnostic performance was significantly better after the application of CAD than it was before, but not in terms of sensitivity (94.1% vs 97.1%; p=.38) or negative predictive value.

Reference:

He, P., Chen, W., Bai, M.-Y., Li, J., Wang, Q.-Q., Fan, L.-H., Zheng, J., Liu, C.-T., Zhang, X.-R., Yuan, X.-R., Song, P.-J., & Cui, L.-G. (2023). Deep Learning–Based Computer-Aided Diagnosis for Breast Lesion Classification on Ultrasound: A Prospective Multicenter Study of Radiologists Without Breast Ultrasound Expertise. In American Journal of Roentgenology. American Roentgen Ray Society. https://doi.org/10.2214/ajr.23.29328