Human and AI collaboration for pulmonary nodule segmentation

Hi-Seg, a human-in-the-loop segmentation framework built on the Segment Anything Model, achieved 85% accuracy in pulmonary nodule detection across 1,179 patients, outperforming five state-of-the-art AI models by 10-22%. The research demonstrates that non-experts with brief training can match junior medical professionals' performance, suggesting foundation models can be safely integrated into clinical workflows while reducing annotator burden.

Hi-Seg represents a pragmatic approach to deploying foundation models in healthcare settings where specialist expertise remains scarce. Rather than replacing human judgment, the framework positions AI as a collaborative tool refined through iterative human feedback, addressing a critical gap between general-purpose AI capabilities and domain-specific medical requirements. The study's external validation across 12 centers and multiple annotator groups—including non-medical personnel—provides robust evidence that human-in-the-loop workflows can democratize medical annotation without compromising quality.

This research responds to a broader tension in medical AI deployment. Foundation models like SAM show remarkable generalization capabilities, yet their application to specialized medical tasks often underperforms domain-specific models without human guidance. Hi-Seg's 10-22% accuracy improvement over existing deep learning approaches demonstrates that human reasoning, even from briefly trained non-experts, introduces semantic understanding that pure neural networks lack. The framework also addresses economic constraints in clinical settings by reducing annotation time while enabling crowdsourced labeling.

For healthcare technology and AI development, this validates an emerging paradigm where foundation models serve as collaborative assistants rather than autonomous decision-makers. The finding that non-medical annotators achieved comparable performance to junior medical students has significant implications for scaling medical AI systems in resource-constrained regions. This approach could accelerate the adoption of AI-assisted diagnostics by lowering expertise barriers and building clinician confidence through transparent, human-guided segmentation processes.

Future work will likely explore Hi-Seg's applicability to other medical imaging tasks and organs, testing whether human-in-the-loop frameworks generalize across pathology types and imaging modalities.

• →Hi-Seg achieved 85% Dice score, outperforming five state-of-the-art models by 10-22% through human-AI collaboration on pulmonary nodule segmentation
• →Non-medical annotators with brief training matched junior medical student performance, suggesting democratized annotation workflows are viable
• →The framework reduced annotation time while improving accuracy, addressing clinician workload concerns in medical practice
• →External validation across 1,179 patients and 12 centers provides robust evidence for safe foundation model integration into clinical workflows
• →Human-in-the-loop segmentation enables scalable crowdsourced annotation and transforms how specialized AI models deploy in healthcare