Image2SSM: Reimagining Statistical Shape Models from Images with Radial Basis Functions.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning in biology

Background:

• Statistical shape modeling (SSM) is crucial for analyzing anatomical variations.
• Traditional SSM pipelines involve complex, time-consuming segmentation and registration steps.
• Existing methods for compact shape representation are often laborious and costly.

Purpose of the Study:

• To introduce Image2SSM, a novel deep-learning approach for automated SSM.
• To leverage image-segmentation pairs for direct learning of Radial-Basis-Function (RBF)-based shape representations.
• To enable scalable SSM for large biological datasets with minimal user intervention.

Main Methods:

• Image2SSM utilizes deep learning to learn an RBF-based shape representation directly from image-segmentation pairs.
• The RBF representation provides a self-supervised signal for estimating continuous, compact surface representations.
• The method adapts to complex geometries in a data-driven manner, facilitating statistical landmark-based shape model construction.

Main Results:

• Image2SSM successfully characterizes populations of biological structures by constructing statistical shape models.
• The approach requires minimal parameter tuning and no user assistance.
• Experiments on synthetic and real datasets demonstrate superior efficacy compared to state-of-the-art correspondence-based methods.

Conclusions:

• Image2SSM offers a scalable and efficient deep-learning-based solution for statistical shape modeling.
• The method automates the extraction of low-dimensional shape representations from unsegmented images.
• Image2SSM has the potential to significantly advance SSM applications, particularly for large-scale biomedical studies.