Prediction of benign and malignant ovarian tumors using Resnet34 on ultrasound images
Kuo Miao, Ning Zhao, Qian Lv, Xin He, Mingda Xu, Xiaoqiu Dong, Dandan Li, Xiaohui Shao
Abstract
Abstract Objective To develop deep learning (DL) prediction models using transvaginal ultrasound (TVS), transabdominal ultrasound (TAS), and color Doppler flow imaging (CDFI) of TVS (CDFI_TVS) to automatically predict benign or malignant ovarian tumors. Methods This retrospective study included women with ovarian tumors who underwent ultrasound between August 2018 and October 2022. Histopathological analysis was used as a reference standard. The dataset was preprocessed by clipping, flipping, and rotating images to generate a larger, more complicated, and diverse dataset to improve accuracy and generalizability. The dataset was then divided into training (80%) and test (20%) sets. The weights of the models, modified from the residual network (ResNet) with the TVS, TAS, and CDFI_TVS images (hereafter, referred to as DL TVS , DL TAS , and DL CDFI_TVS , respectively) were developed. The area under the receiver operating characteristic curve (AUC) analysis in the test set was used to compare the predictive value of DL for malignancy. Results A total of 2340 images from 1350 women with adnexal masses were included. DL TVS had an AUC of 0.95 (95% CI: 0.93–0.97) for classifying malignant and benign ovarian tumors, comparable with that of DL TAS (AUC, 0.95; 95% CI: 0.91–0.98; p = 0.96) and DL CDFI_TVS (AUC, 0.88; 95% CI: 0.84–0.93; p = 0.02). Decision curve analysis indicated that DL TVS performed better than DL TAS and DL CDFI_TVS . Conclusion We developed DL models based on TVS, TAS, and CDFI_TVS on ultrasound images to predict benign and malignant ovarian tumors with high diagnostic performance. The DL TVS model had the best prediction compared with the DL TAS and DL CDFI_TVS models.