Deep learning to identify breast disease in an 87-patient clinical study of breast core biopsies to provide rapid biopsy evaluation

Optical coherence tomography (OCT) is being studied to provide rapid biopsy evaluation. Here we developed a deep learning algorithm to rapidly identify disease in OCT images in an 87-patient IRB-approved clinical study. Pathologists labelled each biopsy into two categories: non-interest (no disease) and interest (for further pathological analysis). Our dataset was split by patients into training (n = 70) and validation (n = 17). The Resnet18 architecture used the Adam optimizer, had a learning rate of 0.01, batch size of 8, and ran for 30 epochs. The network achieved 97% training accuracy and 70% validation accuracy.