Extending pretrained segmentation networks with additional anatomical structures.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Anatomy

Background:

• Comprehensive surgical planning requires accurate segmentation of anatomical structures using patient-specific models.
• Deep neural networks (DNNs) excel at segmentation but require extensive annotated data, which is often scarce and costly.
• Building segmentation models incrementally from diverse datasets is crucial for practical applications.

Purpose of the Study:

• To develop a class-incremental framework for extending deep segmentation networks to new anatomical structures.
• To enable model adaptation using minimal incremental annotations without full retraining.
• To address the challenge of limited annotated data in medical image segmentation.

Main Methods:

• A class-incremental framework was proposed to extend existing deep segmentation networks.
• Knowledge distillation from the current network state was employed to prevent catastrophic forgetting.
• The method utilizes minimal incremental annotation sets for new anatomical structures.

Main Results:

• The proposed method demonstrated minimal Dice score loss (<1%) when retaining performance on old classes with 50% incremental annotations, compared to 25% loss with conventional fine-tuning.
• In a one-shot incremental learning setting, the framework outperformed vanilla network performance by over 11% in Dice score.
• The framework inherently leverages transferable knowledge from previously trained structures.

Conclusions:

• The method allows learning new anatomical structures while preserving performance on older ones, suitable for lifelong class-incremental learning.
• A fraction of annotations is sufficient for incrementally building comprehensive segmentation models by leveraging prior knowledge.
• The framework is adaptable for future network architectures with minor additions per structure.