Benchmark-Ready 3D Anatomical Shape Classification

TL;DR

The paper tackles the lack of standardized benchmarks and robust learning approaches for anatomical 3D shape classification by introducing MedShapeNet19, a curated 15,200-sample benchmark across 19 anatomical classes, and PSPooling, a non-learnable, precomputed pooling operator that preserves geometric structure during mesh coarsening via a fixed correspondence $P \in \mathbb{R}^{|V'| \times |V|}$ defined by ${\mathcal{R}}$, bounded by $k_S$. Using a self-supervised graph autoencoder, the method learns anatomy-aware representations and demonstrates that PSPooling leads to smoother reconstructions, more discriminative latent spaces, and competitive downstream classification, especially in low-label regimes. The experiments on MedShapeNet19 show that PSPooling outperforms traditional SAG pooling in reconstruction fidelity and robustness to limited supervision, with pretrained encoders achieving accuracy on par with, or better than, fully supervised models using far fewer labels. Overall, MedShapeNet19 provides a practical benchmark for fair comparisons, and PSPooling offers a scalable, structure-preserving pooling paradigm for medical 3D shape analysis that can be extended to segmentation and reconstruction.

Abstract

Progress in anatomical 3D shape classification is limited by the complexity of mesh data and the lack of standardized benchmarks, highlighting the need for robust learning methods and reproducible evaluation. We introduce two key steps toward clinically and benchmark-ready anatomical shape classification via self-supervised graph autoencoding. We propose Precomputed Structural Pooling (PSPooling), a non-learnable mesh pooling operator designed for efficient and structure-preserving graph coarsening in 3D anatomical shape analysis. PSPooling precomputes node correspondence sets based on geometric proximity, enabling parallelizable and reversible pooling and unpooling operations with guaranteed support structure. This design avoids the sparsity and reconstruction issues of selection-based methods and the sequential overhead of edge contraction approaches, making it particularly suitable for high-resolution medical meshes. To demonstrate its effectiveness, we integrate PSPooling into a self-supervised graph autoencoder that learns anatomy-aware representations from unlabeled surface meshes. We evaluate the downstream benefits on MedShapeNet19, a new curated benchmark dataset we derive from MedShapeNet, consisting of 19 anatomical classes with standardized training, validation, and test splits. Experiments show that PSPooling significantly improves reconstruction fidelity and classification accuracy in low-label regimes, establishing a strong baseline for medical 3D shape learning. We hope that MedShapeNet19 will serve as a widely adopted benchmark for anatomical shape classification and further research in medical 3D shape analysis. Access the complete codebase, model weights, and dataset information here: https://github.com/TomasKrsicka/MedShapeNet19-PSPooling.

Figures (5)

• Figure 1: MedShapeNet19: a curated benchmark dataset for 3D medical shape classification derived from MedShapeNet Li2024medshapenet. We refine an initially large-scale and weakly curated dataset by filtering erroneous images, removing non-body structures such as instruments, and merging semantically similar labels.
• Figure 2: Method outline of our proposed Precomputed Structural Pooling operator. Node correspondences are precomputed between successive levels of the subsampled meshes. Node A gathers features from nodes 1, 2 and 3. Node B from nodes 4, 5 and 6.
• Figure 3: Comparison of reconstruction quality between architectures utilizing SAG pooling and PSPooling. PSPooling reconstructions show better detail retention, preserve disconnected components and prevent formation of vertex clustering artifacts.
• Figure 4: t-SNE visualizations of latent spaces produced by autoencoders employing SAGPool (left) and PSPooling (right). The latent space from a network that uses PSPooling exhibits more pronounced clustering tendencies.
• Figure 5: Downstream classification accuracy comparison of pooling strategies across various dataset fractions. While both methods show performance degradation with reduced training data, networks using PSPooling (red curves) demonstrate greater robustness under limited supervision, showing a smaller performance drop compared to SAGPool (blue curves).