Real-time intraoperative identification of malignant ovarian tumors using deep learning of clinical gross specimen images.

OBJECTIVE: Ovarian cancer is one of the most common gynecologic malignancies with high mortalities. A proper operation strategy is key to improve patients' prognosis. However, a timely intra-operative identification of malignant ovarian tumor is challenging for clinical physicians. This study aims to develop a deep learning model that can be used intra-operatively to differentiate malignant and benign ovarian tumors based on gross specimen images. MATERIALS AND METHODS: This retrospective cohort study was conducted in an academic tertiary hospital. The study collected 662 cut-in-half gross images of the specimen from 243 patients as a training data set to develop convolution neural network (CNN)-based models. Different deep learning algorithms including classic CNN, VGG16, Xception, InceptionV3, MobileNet, NASNet, EfficientNetB0 and ResNet 50 were used for training and validation. Another 56 patients with 129 cut-in-half images were used as the independent test data set. Smartphone application with the integration of the optimized model was developed to perform intraoperative identification of malignant ovarian tumors. RESULTS: Among the eight different CNN models, the MobileNet model had a better discriminating ability in classification (accuracy: 69.77%, recall: 61.84%, precision: 82.46%, area under curve (AUC): 0.7149) and the performance was more precise compared with junior physicians (accuracy: 71.83%, recall: 80.48%, precision: 74.33%). CONCLUSIONS: The real-time intra-operative assessment of ovarian tumor using the mobile application has the potential to accelerate clinical decision making before pathology reports during the operation.