Ovarian cancer exhibits the highest death rate among gynecological cancers and is typically first evaluated through ultrasound. However, the intricate nature of ultrasound visuals for ovarian lesions, combined with the deep pelvic positioning, demands substantial expertise for accurate subjective interpretation. Consequently, identifying ovaries and ovarian lesions, along with diagnosing ovarian cancer, remains difficult. This study sought to create an automated deep learning-based system called the Ovarian Multi-Task Attention Network (OvaMTA). It targets ovary and ovarian mass detection, segmentation, and classification, while also supporting the diagnosis of ovarian masses from ultrasound screenings. From June 2020 to May 2022, the OvaMTA model underwent training, validation, and testing using a training/validation group of 6938 images and an internal test group of 1584 images. These were gathered from 21 hospitals among women receiving ultrasound for ovarian masses. Two additional external test groups were then assembled from two other hospitals. This included 1896 images collected from February 2024 to April 2024 for the image-based external test set, plus 159 videos gathered from April 2024 to May 2024 for the video-based external test set. An artificial intelligence (AI) platform named OvaMTA was built to diagnose ovarian masses via ultrasound screening. It comprises two components: OvaMTA-Seg, a whole-image segmentation model for ovary detection, and OvaMTA-Diagnosis, a diagnostic model that predicts the pathological nature of ovarian masses from image patches extracted by OvaMTA-Seg. System performance was assessed across one internal and two external validation sets, with comparisons to physicians' evaluations in practical clinical testing. Eight doctors participated in reviewing the real-world cases. The AI's value in supporting physician diagnoses was also examined. For segmentation tasks, OvaMTA-Seg attained an average Dice score of 0.887 on the internal test set and 0.819 on the image-based external test set. It also showed strong results in detecting ovarian masses across test images (covering both normal ovaries and lesions), with an internal test area under the curve (AUC) of 0.970 and an external test AUC of 0.877. Regarding classification and diagnostic prediction, OvaMTA-Diagnosis delivered strong outcomes on the image-based internal (AUC: 0.941) and external (AUC: 0.941) test sets. In the video-based external evaluation, OvaMTA processed 159 videos containing ovarian masses, achieving an AUC of 0.911. This performance was similar to that of senior radiologists (accuracy [ACC]: 86.2% vs. 88.1%, p = 0.50; sensitivity [SEN]: 81.8% vs. 88.6%, p = 0.16; specificity [SPE]: 89.2% vs. 87.6%, p = 0.68). AI assistance led to notable gains for junior and intermediate radiologists compared to their unaided performance (ACC: 80.8% vs. 75.3%, p = 0.00015; SEN: 79.5% vs. 74.6%, p = 0.029; SPE: 81.7% vs. 75.8%, p = 0.0032). General practitioners supported by AI reached performance levels matching those of radiologists on average (ACC: 82.7% vs. 81.8%, p = 0.80; SEN: 84.8% vs. 82.6%, p = 0.72; SPE: 81.2% vs. 81.2%, p > 0.99). The ultrasound-based OvaMTA system serves as a straightforward and effective support tool for ovarian cancer screening, delivering diagnostic accuracy on par with senior radiologists. It represents a promising resource for ovarian cancer screening efforts.
