MRI-based Ovarian Lesion Classification via a Foundation Segmentation Model and Multimodal Analysis: A Multicenter Study.

Background Artificial intelligence may enhance diagnostic accuracy in classifying ovarian lesions on MRI scans; however, its applicability across diverse datasets is uncertain. Purpose To develop an efficient, generalizable pipeline for MRI-based ovarian lesion characterization. Materials and Methods In this retrospective study, multiparametric MRI datasets of patients with ovarian lesions from a primary institution (January 2008 to January 2019) and two external institutions (January 2010 to October 2020) were analyzed. Lesions were automatically segmented using Meta's Segment Anything Model (SAM). A DenseNet-121 deep learning (DL) model incorporating both imaging and clinical data was then trained and validated externally for ovarian lesion classification. Lesions were evaluated by radiologists using the Ovarian-Adnexal Reporting and Data System for MRI and subjective assessment, classifying them as benign or malignant. The classification performances of the DL model and radiologists were compared using the DeLong test. Results The primary dataset included 534 lesions from 448 women (mean age, 52 years ± 15 [SD]) from institution A (United States), whereas the external datasets included 58 lesions from 55 women (mean age, 51 years ± 19) from institution B (United States) and 29 lesions from 29 women (mean age, 49 years ± 10) from institution C (Taiwan). SAM-assisted segmentation had a Dice coefficient of 0.86-0.88, reducing the processing time per lesion by 4 minutes compared with manual segmentation. The DL classification model achieved an area under the receiver operating characteristic curve (AUC) of 0.85 (95% CI: 0.85, 0.85) on the internal test and 0.79 (95% CI: 0.79, 0.79 and 0.78, 0.79) across both external datasets with SAM-segmented images, comparable with the radiologists' performance (AUC: 0.84-0.93; all