OP06.02: Deep learning radiomics based on contrast‐enhanced ultrasound images for accurate diagnosis of adnexal lesions: a retrospective multicentre study

Accurate discrimination between malignant and benign adnexal lesions could help avoid unnecessary surgery and overtreatment, and preserving fertility. This study aimed to develop a deep learning radiomics (DLR) model based on contrast-enhanced ultrasound (CEUS) images to assist radiologists for the automated classification of benign versus malignant adnexal lesions. This retrospective, multicentre study included consecutive women with adnexal lesions undergoing CEUS examinations from 14 hospitals across China between January 2018 and September 2023. Histopathologic analysis served as the reference standard. A DLR model based on CEUS images was developed for the characterisation of malignant and benign adnexal lesions. Two rounds of reader studies were performed. The accuracy of the developed model was assessed with receiver operator characteristic, calibration, and decision curve analysis. A total of 405 adnexal lesions with 257 benign and 148 malignant adnexal lesions were retrospectively enrolled. The cohort was split into the training set (n = 280), internal validation set (n = 60) and external validation set (n = 65). The DLR model yielded an area under curve (AUC) of 0.986 (95% CI:0.979-0.993), 0.966 (95% CI:0.957-0.975), 0.928 (95% CI:0.913-0.943) in the training set, internal validation set and external validation set, respectively. With the aid of the DLR model, higher performances were achieved compared to the assessment by the four radiologists alone (0.962 vs 0.914; 0.963 vs 0.927; 0.929 vs 0.835; 0.923 vs 0.830, respectively). The developed DLR model based on CEUS images in our study may serve as an effective and emerging tool to assist radiologists for distinguishing malignant from benign adnexal lesions.