A multimodal uncertainty-aware AI system optimizes ovarian cancer risk assessment workflow
Xiaodong Wang, Xiaohui Lv, Jingwen Wang, Lulu Zou, Zijun Chen, Ruiyu Zhao, Lisha Zhao, Min Zhao, Xinlei Zhang, Boan Zhang, Jiahao Zhang, Yiteng Zhu, Xin Shi, Yane Gao, M. Liu, Lirong Ai, Liming Wang, Xiyang Liu, Hong Yang
Abstract
Accurate ovarian cancer screening and diagnosis are critical for patient survival. We present UMORSS, an AI-assisted diagnostic system integrating ultrasound (US) imaging and clinical data with uncertainty quantification for precise ovarian cancer risk assessment. Developed and evaluated using a multicentre dataset (7352 patients, 7594 lesions, 9281 US images), UMORSS employs a two-phase approach: Phase I rapidly triages low-risk lesions via initial US analysis, and Phase II provides uncertainty-aware multimodal analysis for complex cases. Phase I accurately identified 68.7% of physiological cysts and 13.8% of benign tumours as low-risk, with zero false negatives, and Phase II achieved an AUC of 0.955 (internal testing) and 0.926 (external validation). Furthermore, a prospective reader study (n = 284 cases, six radiologists) demonstrated that UMORSS as a human-AI collaborative tool increased radiologists' average AUC by 10.58% and sensitivity by 22.48%. UMORSS shows strong potential to streamline clinical workflow, optimize resource allocation, and standardize ovarian cancer diagnosis.