Point2SSM++: Self-Supervised Learning of Anatomical Shape Models from Point Clouds

TL;DR

Point2SSM++ tackles the preprocessing and bias issues that limit the adoption of correspondence-based SSM in clinical settings by learning anatomical point correspondences directly from unaligned point clouds in a self-supervised fashion. It introduces misalignment-robust input handling, a consistency loss to enforce stable correspondences across samplings and poses, and extensions for multi-anatomy and 4D spatiotemporal data using PSTNet2. The approach demonstrates superior or competitive performance against state-of-the-art deep learning baselines and traditional PSM across diverse anatomies, while enabling downstream tasks such as disease classification and longitudinal studies. By removing strict input prerequisites and reducing computational overhead, Point2SSM++ broadens the practical utility of SSM for clinical research and patient-specific analyses.

Abstract

Correspondence-based statistical shape modeling (SSM) stands as a powerful technology for morphometric analysis in clinical research. SSM facilitates population-level characterization and quantification of anatomical shapes such as bones and organs, aiding in pathology and disease diagnostics and treatment planning. Despite its potential, SSM remains under-utilized in medical research due to the significant overhead associated with automatic construction methods, which demand complete, aligned shape surface representations. Additionally, optimization-based techniques rely on bias-inducing assumptions or templates and have prolonged inference times as the entire cohort is simultaneously optimized. To overcome these challenges, we introduce Point2SSM++, a principled, self-supervised deep learning approach that directly learns correspondence points from point cloud representations of anatomical shapes. Point2SSM++ is robust to misaligned and inconsistent input, providing SSM that accurately samples individual shape surfaces while effectively capturing population-level statistics. Additionally, we present principled extensions of Point2SSM++ to adapt it for dynamic spatiotemporal and multi-anatomy use cases, demonstrating the broad versatility of the Point2SSM++ framework. Furthermore, we present extensions of Point2SSM++ tailored for dynamic spatiotemporal and multi-anatomy scenarios, showcasing the broad versatility of the framework. Through extensive validation across diverse anatomies, evaluation metrics, and clinically relevant downstream tasks, we demonstrate Point2SSM++'s superiority over existing state-of-the-art deep learning models and traditional approaches. Point2SSM++ substantially enhances the feasibility of SSM generation and significantly broadens its array of potential clinical applications.

Figures (20)

• Figure 1: A. Point2SSM adams2023point2ssm architecture B. Point2SSM++ architecture C. Point2SSM++, Classifer architecture D. PSTNet fan2021pstnet2 4D encoder architecture.
• Figure 2: Mode of variation in half torus bump shapes.
• Figure 3: Results on the nonlinear half torus bump dataset from PSM and Point2SSM++. The three example test predictions (top) and resulting primary and secondary modes of variation (bottom) demonstrate that Point2SSM++ correctly captures nonlinear variation, whereas PSM fails to do so.
• Figure 4: Single anatomy experiment results. Boxplots show the error distribution across test sets for each model in mm. Compactness plots show the cumulative variance ratio over the number of PCA modes. Generalization and specificity plots show the CD over the number of PCA modes. A maximum of 30 modes is displayed which captures at least $99\%$ of the total variation for all datasets. Lower values are better for all metrics except PCA compactness.
• Figure 5: Primary and secondary modes of population variation captured by the PSM (gray) and Point2SSM++ (green) methods across each dataset. The mean and shapes at $\pm 1.5$ standard deviation are shown with color denoting the distance to the mean. Blue indicates the surface is outside the mean, and red indicates the surface is inside the mean. Separate color scales and orientation markers are provided for each dataset.