Automated Ovarian Cancer Identification Using End-to-End Deep Learning and Second Harmonic Generation Imaging

Abstract

Surgery is one of the most important methods for the treatment of ovarian cancer. During this procedure, a biopsy is usually required to evaluate the suspicious lesions and provide guidance for the size of the surgical resection. However, the conventional biopsy for intraoperative histopathological diagnosis performed by trained pathologists is labor-intensive, time-consuming, and carries the risk of bias. Therefore, we present a novel optical biopsy method to assist physicians in accurately and rapidly diagnosing ovarian cancer during surgery. We demonstrate that second harmonic generation (SHG) images of unstained, freshly resected ovarian tissues can be accurately characterized by deep learning techniques. Using 13,563 SHG images obtained from freshly resected human ovarian tissues of 74 patients, we fine-trained a convolutional neural network (CNN) based on pretrained ResNet50 framework to distinguish normal, benign, and malignant ovarian tissue with an average accuracy of 99.7%. These results suggest that optical biopsies based on label-free SHG imaging and deep learning technology have great potential for rapid and accurate characterizations of ovarian lesions in surgery.