Abstract
Medical image segmentation is vital for precise identification and analysis of anatomical structures and pathological regions, yet traditional models often fall short in aligning with clinical workflows, requiring extensive manual correction even when overall segmentation accuracy is high. To address this gap, we introduce HybridMS, a hybrid intelligence framework designed to maintain high segmentation accuracy while substantially reducing clinician workload through selective human intervention. HybridMS employs an uncertainty-driven feedback mechanism that selectively triggers clinician input only for cases predicted to be challenging, thereby avoiding unnecessary manual review. Corrected cases are prioritised during retraining through a weighted update strategy, enabling the model to adapt more effectively to clinically relevant errors. This design minimises intervention frequency while preserving segmentation quality. Evaluated on lung segmentation in chest X-rays for tuberculosis detection, HybridMS achieved comparable or improved performance over the baseline MedSAM model (Dice: 0.9538 vs. 0.9435; IoU: 0.9126 vs. 0.8941) with consistent boundary quality in difficult cases. For the subset of cases identified as challenging (baseline Dice < 0.92), HybridMS reduced mean Hausdorff Distance and Average Symmetric Surface Distance, demonstrating more stable anatomical boundaries. Workflow efficiency was markedly improved: in a preliminary timing study with radiologists, average annotation time was reduced by approximately 82% for standard cases and 60% for challenging cases, without compromising accuracy. By combining targeted human oversight with automated refinement, HybridMS demonstrates that stable segmentation performance can be achieved with significantly lower annotation effort, offering a clinically viable pathway for efficient and reliable deployment in diagnostic workflows.