Breast cancer classification in point-of-care ultrasound imaging-the impact of training data.

Abstract

Purpose: The survival rate of breast cancer for women in low- and middle-income countries is poor compared with that in high-income countries. Point-of-care ultrasound (POCUS) combined with deep learning could potentially be a suitable solution enabling early detection of breast cancer. We aim to improve a classification network dedicated to classifying POCUS images by comparing different techniques for increasing the amount of training data.

Approach: Two data sets consisting of breast tissue images were collected, one captured with POCUS and another with standard ultrasound (US). The data sets were expanded by using different techniques, including augmentation, histogram matching, histogram equalization, and cycle-consistent adversarial networks (CycleGANs). A classification network was trained on different combinations of the original and expanded data sets. Different types of augmentation were investigated and two different CycleGAN approaches were implemented.

Results: Almost all methods for expanding the data sets significantly improved the classification results compared with solely using POCUS images during the training of the classification network. When training the classification network on POCUS and CycleGAN-generated POCUS images, it was possible to achieve an area under the receiver operating characteristic curve of 95.3% (95% confidence interval 93.4% to 97.0%).

Conclusions: Applying augmentation during training showed to be important and increased the performance of the classification network. Adding more data also increased the performance, but using standard US images or CycleGAN-generated POCUS images gave similar results.