Mesh2SSM++: A Probabilistic Framework for Unsupervised Learning of Statistical Shape Model of Anatomies from Surface Meshes

• DOI: 10.48550/arxiv.2502.07145
• Publication Date: 2025-02-10
• Authors: Krithika Iyer; Mokshagna Sai Tej...; Shireen Elhabian
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Anatomy evaluation is crucial for understanding the physiological state, diagnosing abnormalities, and guiding medical interventions. Statistical shape modeling (SSM) vital in this process. By enabling extraction of quantitative morphological descriptors from MRI CT scans, SSM provides comprehensive descriptions anatomical variations within a population. However, effectiveness anatomy hinges on quality robustness models. While deep learning techniques show promise addressing these challenges by complex nonlinear representations shapes, existing models still have limitations often require pre-established training. To overcome issues, we propose Mesh2SSM++, novel approach that learns to estimate correspondences meshes an unsupervised manner. This method leverages unsupervised, permutation-invariant representation how deform template point cloud into subject-specific meshes, forming correspondence-based model. Additionally, our probabilistic formulation allows population-specific template, reducing potential biases associated with selection. A key feature Mesh2SSM++ its ability quantify aleatoric uncertainty, which captures inherent data variability essential ensuring reliable model predictions robust decision-making clinical tasks, especially under challenging imaging conditions. Through extensive validation across diverse anatomies, metrics, downstream demonstrate outperforms methods. Its operate directly combined computational efficiency interpretability through framework, makes it attractive alternative traditional learning-based approaches.