A multimodal uncertainty-aware AI system optimizes ovarian cancer risk assessment workflow

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate ovarian cancer screening and diagnosis are crucial for improving patient survival rates.
• Current diagnostic methods can be enhanced with advanced AI tools for better risk assessment.

Purpose of the Study:

• To develop and evaluate UMORSS, an AI-assisted diagnostic system for precise ovarian cancer risk assessment.
• To integrate ultrasound imaging and clinical data with uncertainty quantification for improved diagnostic accuracy.

Main Methods:

• UMORSS utilizes a two-phase approach: Phase I for rapid triage of low-risk lesions and Phase II for uncertainty-aware multimodal analysis.
• The system was developed and validated on a multicentre dataset comprising 7352 patients, 7594 lesions, and 9281 ultrasound images.
• A prospective reader study involving six radiologists assessed UMORSS as a human-AI collaborative tool.

Main Results:

• Phase I correctly identified low-risk lesions with zero false negatives.
• Phase II achieved an AUC of 0.955 (internal) and 0.926 (external validation).
• The human-AI collaboration with UMORSS increased radiologists' average AUC by 10.58% and sensitivity by 22.48%.

Conclusions:

• UMORSS demonstrates significant potential in streamlining clinical workflows for ovarian cancer diagnosis.
• The AI system can optimize resource allocation and standardize diagnostic processes.
• UMORSS enhances diagnostic accuracy and supports radiologists in complex cases, improving overall patient care.