Generalizable multi-task, multi-domain deep segmentation of sparse pediatric imaging datasets via multi-scale...

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Accurate diagnosis of pediatric musculoskeletal disorders relies on medical imaging analysis.
• Deep learning-based semantic segmentation generates 3D anatomical models for morphological evaluation.
• Limited pediatric imaging data hinders the accuracy and generalization of segmentation models.

Purpose of the Study:

• To develop a novel multi-task, multi-domain deep learning framework for pediatric musculoskeletal imaging.
• To address data scarcity by leveraging multiple datasets and domains within a single network.
• To enhance the accuracy and generalization of automated anatomical model generation.

Main Methods:

• Implemented a multi-task, multi-domain learning framework optimizing a single network over diverse pediatric imaging datasets.
• Incorporated transfer learning from natural image classification and multi-scale contrastive regularization.
• Utilized multi-joint anatomical priors to ensure anatomically consistent segmentation predictions.
• Evaluated bone segmentation on scarce pediatric datasets of the ankle, knee, and shoulder joints.

Main Results:

• The proposed multi-task, multi-domain framework significantly outperformed individual, transfer, and shared segmentation models.
• Achieved statistically significant improvements in Dice metric for pediatric bone segmentation.
• Demonstrated enhanced generalization capabilities by effectively utilizing limited and diverse imaging resources.

Conclusions:

• The novel framework offers a promising solution for overcoming data scarcity in pediatric medical image segmentation.
• This approach facilitates the intelligent use of imaging resources for improved diagnosis and management of pediatric musculoskeletal disorders.
• The study highlights the potential of multi-domain learning in advancing medical image analysis for pediatric populations.