Development of a Deep-Learning Algorithm for Detecting Suspicious Breast Lesions on Chest CT

A convolutional neural network (CNN) was trained and evaluated for detecting suspicious breast lesions on a large dataset of chest CT exams from a teleradiology practice covering over 2,000 hospital sites. Radiologists annotated any discrete nodules or masses appearing within breast tissue, and the model was tested on a held-out set. At a threshold achieving 0.99 specificity, the model demonstrated a sensitivity of 0.32 and a positive predictive value (PPV) of 0.50. In a scenario where sensitivity and specificity were balanced, the model achieved sensitivity and specificity of 0.79 each, resulting in a PPV of 0.10. The overall performance, as indicated by the area under the receiver operating characteristic curve (AUC), was 0.87. These results highlight the potential of an automated system to identify suspicious breast lesions on chest CT exams, thereby aiding in the opportunistic detection of malignancies.