Enhanced ovarian cancer diagnosis using deep learning on pelvic ultrasound with integrated clinical data: retrospective multicenter study

초록

OBJECTIVE: This study aimed to develop a deep learning (DL) model to enhance the differential diagnosis of benign and malignant ovarian tumors by integrating pelvic ultrasound images with clinical data, such as age and cancer antigen (CA)-125 levels. METHODS: We analyzed pelvic ultrasound images, age, and CA-125 levels from 804 patients diagnosed with ovarian tumors (446 benign, 358 malignant) from 2015 to 2022. Images were segmented into cystic and solid components for feature extraction. Patients were divided into training (n=565), validation (n=76), and test datasets (n=163). ResNet50 and DenseNet121 models were trained on these data, with clinical information added to classifier architecture for improved prediction accuracy. RESULTS: Using ultrasound images alone, ResNet50 and DenseNet121 achieved areas under the receiver operating characteristic curve (AUCs) of 0.84 and 0.82, respectively. When clinical data and segmented solid images were included, AUCs improved to 0.95 for ResNet50 and 0.96 for DenseNet121. For the test set, ResNet50 and DenseNet121 achieved sensitivities of 90% and 81%, specificities of 93% and 97%, positive predictive values of 92% and 95%, and negative predictive values of 92% and 86%. CONCLUSIONS: Binary classification model based on DL algorithms using ultrasound images can distinguish between benign and malignant ovarian tumors accurately. Segmentation of solid portion and clinical information of age and CA-125 at diagnosis combined with the pelvic ultrasound images increased the accuracy of the classification model. © 2026. Asian Society of Gynecologic Oncology, Korean Society of Gynecologic Oncology, and Japan Society of Gynecologic Oncology.

키워드