Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle

Area of Science:

• Medical image analysis
• Computer-aided diagnosis
• Machine learning for healthcare

Background:

• Medical image segmentation often involves ambiguous tissues/lesions, requiring multiple expert annotations to reduce bias.
• Current annotation fusion methods like majority vote overlook grader expertise differences.
• Medical image segmentation primarily aids clinical disease diagnosis.

Purpose of the Study:

• To propose a novel framework, the Diagnosis-First segmentation Framework (DiFF), to calibrate inter-observer segmentation uncertainty.
• To leverage disease diagnosis as the primary criterion for refining segmentation accuracy.
• To develop a method that enhances segmentation performance by considering expert differences.

Main Methods:

• DiFF first learns to fuse multi-rater segmentation labels into a single ground-truth (Diagnosis-First Ground-truth, DF-GT) optimized for disease diagnosis.
• A Take and Give Model (T&G Model) is proposed to segment the DF-GT from raw medical images.
• The framework integrates calibrated uncertainty into the segmentation process to aid diagnosis.

Main Results:

• DiFF was validated on optic-disc/optic-cup, thyroid nodule, and skin lesion segmentation tasks.
• The framework effectively calibrated segmentation uncertainty, significantly improving disease diagnosis.
• DiFF demonstrated superior performance compared to existing state-of-the-art multi-rater learning methods.

Conclusions:

• The proposed DiFF framework successfully calibrates segmentation uncertainty by prioritizing disease diagnosis.
• This approach enhances the utility of medical image segmentation for clinical decision-making.
• DiFF offers a significant advancement in multi-rater learning for medical image analysis and diagnosis.