Hybrid intelligence in medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Medical image segmentation is crucial for diagnostics but often requires extensive manual correction, hindering clinical workflow efficiency.
• Existing models struggle to balance segmentation accuracy with reduced clinician intervention needs.

Purpose of the Study:

• To introduce HybridMS, a hybrid intelligence framework designed to improve medical image segmentation accuracy and reduce clinician workload.
• To develop an uncertainty-driven feedback mechanism for selective human intervention in segmentation tasks.
• To enhance model adaptability through prioritized retraining on clinically relevant errors.

Main Methods:

• HybridMS utilizes an uncertainty-driven feedback mechanism to identify and request clinician input only for challenging segmentation cases.
• A weighted update strategy prioritizes corrected cases during model retraining for adaptive learning.
• The framework was evaluated on lung segmentation in chest X-rays for tuberculosis detection, comparing performance against the MedSAM baseline.

Main Results:

• HybridMS achieved comparable or superior segmentation performance (Dice, IoU) to the baseline MedSAM model.
• The framework demonstrated improved boundary quality and reduced surface distance metrics in challenging cases.
• Radiologist annotation time was reduced by approximately 82% for standard and 60% for challenging cases, without compromising accuracy.

Conclusions:

• HybridMS offers a clinically viable solution for efficient and reliable medical image segmentation by combining targeted human oversight with automated refinement.
• The framework significantly lowers annotation effort while maintaining stable segmentation performance.
• HybridMS presents a pathway for integrating AI-driven segmentation into diagnostic workflows with reduced clinician burden.