Automated ultrasound computed tomography for breast mass screening: A multicenter study

Background The low sensitivity of mammography creates a critical screening gap for women with dense breasts in Asia, for whom conventional two-dimensional ultrasonography is often used. However, this method is hampered by significant operator dependence and poor reproducibility. This study aimed to develop an automated breast mass screening tool using artificial intelligence (AI)-aided ultrasound computed tomography (USCT). Methods This retrospective multicenter study analyzed data from July 2022 to March 2024 of 514 patients divided into a development set (n = 222) and two independent external test sets (n = 218; n = 74). Reference standards were established based on clinical ultrasound reports. A multiple instance learning (MIL) model for binary mass screening classification was developed and evaluated via cross-validation and on external test sets using area under the curve (AUROC), with sensitivity and specificity indicated as Youden's index. Results A total of 506 women (mean age ± standard deviation, 45 ± 12 years) contributing 941 breasts were evaluated. In internal cross-validation, the model for breast-level classification achieved a mean AUROC of 0.94 ± 0.02. This performance generalized well to the independent external test set (216 women; 419 breasts), where the model attained an AUROC of 0.91 (95% confidence interval [CI], 0.89–0.94) with a sensitivity of 81% (251 of 308 breasts) and specificity of 89% (99 of 111 breasts) in breast mass screening. This generalizability was further confirmed on a second external test set (73 women; 102 breasts), where the model achieved an AUROC of 0.87 (95% CI, 0.80–0.93) with a sensitivity of 82% (53 of 65 breasts) and specificity of 84% (31 of 37 breasts). Conclusion An AI-aided USCT model was developed that can serve as an operator-independent solution for large-scale breast mass screening in women with dense breasts.