MASSM: An End-to-End Deep Learning Framework for Multi-Anatomy Statistical Shape Modeling Directly From Images

Area of Science:

• Medical Imaging
• Computational Anatomy
• Deep Learning

Background:

• Statistical Shape Modeling (SSM) analyzes anatomical variations but requires manual segmentation, limiting its efficiency.
• Current deep learning methods for SSM often need manual prealignment and bounding box specification.
• Existing approaches struggle with multiple anatomies and direct delineation in image space.

Purpose of the Study:

• To introduce MASSM, a novel end-to-end deep learning framework for automated SSM.
• To enable simultaneous localization, statistical representation estimation, and delineation of multiple anatomies.
• To overcome the limitations of manual segmentation and partially manual inference processes.

Main Methods:

• Developed a multitask deep learning network (MASSM) for simultaneous anatomy localization and shape representation.
• Trained the framework on unsegmented images to generate population-level statistical representations.
• Enabled direct delineation of shape representations in image space for multiple anatomies.

Main Results:

• MASSM successfully automates localization, statistical representation estimation, and delineation of multiple anatomies.
• The framework provides superior shape information compared to traditional segmentation networks.
• MASSM demonstrates more accurate and comprehensive shape representations than pixel-wise segmentation.

Conclusions:

• MASSM offers a fully automated solution for Statistical Shape Modeling, eliminating the need for manual segmentation.
• The multitask approach effectively handles multiple anatomies, enhancing shape analysis capabilities.
• MASSM represents a significant advancement over traditional segmentation methods for medical imaging tasks.