Deep Learning-based FemNet for Key Frame Selection in Ultrasound Videos of Breast Cancer Screening: Distilling Responsible Frames with Feature Entropy for Improved Diagnostic Accuracy

Objective: Breast cancer is the leading cause of cancer-related death among women worldwide. However, heavy workload and a shortage of ultrasound specialists hinder the effectiveness of breast cancer screening. In this study, we aimed to develop a novel deep learning-based framework, called FemNet, to automatically select responsible frames from breast ultrasound videos and classify breast nodules. Methods We designed a feature entropy minimization (FEM) technique and integrated it with a deep learning architecture to create the FemNet framework. We used a dataset of 13,702 breast ultrasound images and 2,141 videos to train and evaluate the proposed framework. We conducted a five-fold cross-validation on the videos set to compare the diagnostic performance of FemNet-selected responsible frames with physician-selected ones. We also compared FemNet's performance with that of physicians under different sensitivity and specificity levels based on the Physician-BIRADS system. Results Our results showed that FemNet-selected responsible frames had statistically superior diagnostic performance compared to physician-selected ones, with an area under the receiver operating characteristic curve of 0.916 ± 0.008 vs. 0.906 ± 0.019 (p = 0.012). Moreover, compared to physicians, FemNet achieved a 7.14% improvement in specificity under the Physician-BIRADS's sensitivity level and a 4.27% improvement in sensitivity under the Physician-BIRADS's specificity level. Conclusions Our study demonstrates that FemNet can effectively select a few frames from lengthy ultrasound videos for breast nodule assessment, similarly to physicians. The proposed framework has the potential to reduce the workload of sonographers and empower physicians to improve the accuracy of breast cancer screening.