Tailored multi-organ segmentation with model adaptation and ensemble.

Area of Science:

• Medical Image Analysis
• Deep Learning in Healthcare

Background:

• Multi-organ segmentation is crucial for medical image analysis but limited annotations hinder deep learning model training.
• Existing deep learning methods for multi-organ segmentation often require substantial, labor-intensive annotated datasets.

Purpose of the Study:

• To develop a multi-organ segmentation model that reduces reliance on annotated data for the target dataset.
• To leverage readily available single-organ segmentation models for improved multi-organ segmentation performance.

Main Methods:

• A novel dual-stage approach comprising a Model Adaptation stage and a Model Ensemble stage.
• The Model Adaptation stage enhances the generalization of off-the-shelf single-organ segmentation models to the target domain.
• The Model Ensemble stage distills and integrates knowledge from multiple adapted single-organ models.

Main Results:

• The proposed method effectively utilizes off-the-shelf single-organ segmentation models.
• A tailored multi-organ segmentation model was developed with high accuracy on four abdomen datasets.
• The approach successfully mitigates the challenge of limited annotated data for multi-organ segmentation.

Conclusions:

• The dual-stage method offers an effective solution for multi-organ segmentation using pre-existing single-organ models.
• This approach significantly improves the accuracy and applicability of deep learning in medical image analysis, especially with limited annotations.
• The findings demonstrate a practical way to build accurate multi-organ segmentation models without extensive manual annotation efforts.