From performance to practice: knowledge-distilled segmentator for on-premises clinical workflows.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Deploying high-capacity medical image segmentation models in clinical settings is challenging due to infrastructure limitations and security policies.
• Computational demands of accurate models hinder practical deployment and maintainability in on-premises hospital environments.

Purpose of the Study:

• To develop a deployment-oriented framework using knowledge distillation to create scalable, compact medical image segmentation models.
• To enable the translation of high-performing segmentation models into efficient student models without altering existing inference pipelines.

Main Methods:

• A knowledge distillation framework was developed to train compact student models from larger teacher models.
• The framework was evaluated on the nnU-Net architecture, with additional validation on transformer and heterogeneous teacher-student models.
• Experiments were conducted on multi-site brain MRI and abdominal CT datasets to assess performance and generalizability.

Main Results:

• The distilled student model achieved 98.7% of the teacher's segmentation accuracy under 94% parameter reduction.
• CPU inference latency was reduced by up to 67% without increasing deployment overhead.
• The approach demonstrated cross-modality generalizability on abdominal CT scans.

Conclusions:

• Knowledge distillation offers a practical and reliable method for converting research-grade segmentation models into deployment-ready components.
• The framework facilitates the integration of efficient, maintainable segmentation models into on-premises clinical workflows.
• This approach addresses the constraints of hospital infrastructure, enabling wider adoption of advanced AI in healthcare.