Abstract:Particle-based shape modeling (PSM) is a popular approach to automatically quantify shape variability in populations of anatomies. The PSM family of methods employs optimization to automatically populate a dense set of corresponding particles (as pseudo landmarks) on 3D surfaces to allow subsequent shape analysis. A recent deep learning approach leverages implicit radial basis function representations of shapes to better adapt to the underlying complex geometry of anatomies. Here, we propose an adaptation of this method using a traditional optimization approach that allows more precise control over the desired characteristics of models by leveraging both an eigenshape and a correspondence loss. Furthermore, the proposed approach avoids using a black-box model and allows more freedom for particles to navigate the underlying surfaces, yielding more informative statistical models. We demonstrate the efficacy of the proposed approach to state-of-the-art methods on two real datasets and justify our choice of losses empirically.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: How to represent and analyze the shape variability of anatomical structures more accurately in the construction of Statistical Shape Models (SSM), especially when dealing with complex geometric shapes. Specifically, the author proposes an optimization - based method, using Radial Basis Function (RBF) shape representation to improve the existing Particle Shape Modeling (PSM) method. ### Main Problems and Solutions 1. **Limitations of Existing Methods**: - **Deep Learning Methods**: Although using implicit RBF shape representation can better adapt to complex geometric structures, deep learning methods are usually black - box models, and it is difficult for users to understand and control their processes, and the computational cost is high. - **Traditional PSM Methods**: Traditional PSM methods provide uniform surface sampling and cannot adapt well to the geometric complexity at different scales, especially when representing complex shapes with local or curled features. 2. **Proposed Solutions**: - **Non - Deep Optimization Method**: The author proposes a non - deep optimization method, which combines the advantages of RBF shape representation and can adaptively adjust the control point distribution without relying on deep neural networks to better capture shape details. - **Loss Function Design**: Four loss functions (surface loss, correspondence loss, feature shape loss, and sampling loss) are introduced to ensure that the control points can be accurately distributed on the surface and maintain the correspondence between shapes, while emphasizing the main shape variation patterns. ### Specific Objectives - **Improve Model Accuracy**: By optimizing the positions of control points, the model can represent individual shapes more accurately, especially complex geometric structures. - **Enhance Model Compactness**: Through feature shape loss, the control points are aligned to the main variation patterns, thereby improving the compactness of the model. - **Ensure Correspondence Consistency**: Through correspondence loss, ensure that the control point configurations between different shapes are consistent, thereby improving the reliability of the statistical model. - **Reduce the Number of Required Control Points**: By adaptively adjusting the control point distribution, reduce the number of control points required to represent complex shapes and improve computational efficiency. ### Experimental Verification The author verifies the effectiveness of the proposed method through two real - world datasets (femoral CT scans and left atria). The results show that this method is superior to the existing Image2SSM and PSM methods in terms of correspondence metrics and bidirectional surface distances. In summary, this paper aims to improve the representation ability and analysis accuracy of statistical shape models in complex geometric structures through improved optimization methods and loss function design.

Optimization-Driven Statistical Models of Anatomies using Radial Basis Function Shape Representation

Active shape model based on sparse representation

Particle-Based Shape Modeling for Arbitrary Regions-of-Interest

Ergonomics-inspired Reshaping and Exploration of Collections of Models

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

Optimized Statistical Deformable Surface Models with Manifold Embedding

DeepSSM: A Deep Learning Framework for Statistical Shape Modeling from Raw Images

Anatomically Parameterized Statistical Shape Model: Explaining Morphometry through Statistical Learning

Mesh2SSM: From Surface Meshes to Statistical Shape Models of Anatomy

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

Discriminative and Generative Models for Anatomical Shape Analysison Point Clouds with Deep Neural Networks

Learning Population-level Shape Statistics and Anatomy Segmentation From Images: A Joint Deep Learning Model

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

Benchmarking off-the-shelf statistical shape modeling tools in clinical applications

Optimal Deformable Surface Models for 3D Medical Image Analysis

BOSS: Bones, Organs and Skin Shape Model

Elastic Statistical Shape Analysis of Biological Structures with Case Studies: A Tutorial

Subject Specific Shape Modeling with Incremental Mixture Models

Deep Optimized Priors for 3D Shape Modeling and Reconstruction

From Images to Probabilistic Anatomical Shapes: A Deep Variational Bottleneck Approach

Use of A Dense Surface Point Distribution Model in A Three-Stage Anatomical Shape Reconstruction from Sparse Information for Computer Assisted Orthopaedic Surgery: A Preliminary Study