learn anatomical priors

This work addresses the challenge of high-fidelity generation of anatomical structures, which exhibit complex geometries and diverse topologies. To this end, we propose an implicit diffusion-based generative framework that incorporates skeletal priors. Our method introduces differentiable skeletonization into medical shape generation for the first time, leveraging a differentiable skeleton module to extract global topological structure and fuse it with local surface features. Diffusion modeling is performed in the signed distance function (SDF) implicit space, followed by a neural implicit decoder to produce high-quality shapes. We contribute MedSDF, a large-scale, multi-category medical shape dataset, and demonstrate superior generation and reconstruction performance over existing methods on both MedSDF and a vascular dataset, while maintaining higher computational efficiency.

This work addresses the limitations of traditional statistical shape modeling, which relies on dense annotations and fixed latent representations, thereby struggling to flexibly capture complex anatomical variations. The authors propose MorphoFlow, a framework that learns compact probabilistic shape representations from only sparse surface annotations. MorphoFlow integrates neural implicit representations, a self-decoder architecture, and autoregressive normalizing flows, augmented with an adaptive latent correlation weighting mechanism. This mechanism leverages a sparsity-inducing prior to automatically modulate the contribution of each latent dimension to anatomical variability, eliminating the need for manual hyperparameter tuning. The method enables high-resolution 3D shape generation and uncertainty quantification. Evaluated on lumbar spine and femur datasets, MorphoFlow achieves high-fidelity reconstructions and accurately recovers population-consistent, structured patterns of anatomical variation.

In virtual imaging trials, generating anatomically accurate, clinically relevant patient-specific phantoms with controllable population-level anatomical variations remains challenging. Method: We propose the first implicit neural representation framework for editable anatomical modeling, integrating geometry-prior-guided implicit surface reconstruction, disentangled latent space learning, and topology-adaptive deformation—enabling fine-grained, target-specific morphological editing of topologically variable organs (e.g., thyroid). Contribution/Results: Our approach is the first to achieve explicit shape–topology disentanglement in anatomical implicit neural representations, supporting clinically interpretable, parameterized editing. Quantitative and qualitative evaluations demonstrate state-of-the-art performance in reconstruction accuracy and anatomical plausibility. Generated phantoms exhibit high fidelity, clinical interpretability, and strong controllability—facilitating reproducible, patient-population-aware virtual imaging studies.

This paper addresses the challenge of locally controlling geometric and topological attributes in multi-class anatomical voxel map generation. Methodologically, it proposes a controllable latent diffusion framework featuring: (1) cubic control domains for substructure-level spatial localization; (2) the first integration of persistent homology into the diffusion process, enabling differentiable topological loss functions to precisely constrain global features—including connected components, cycles, and voids; and (3) voxel moment guidance for geometric shape control, coupled with a neural field decoder for efficient local editing. The approach jointly optimizes multi-dimensional geometric–topological constraints under arbitrary coordinate systems, preserving anatomical plausibility while significantly enhancing generation controllability and interpretability. Experimental results demonstrate superior fidelity and structural compliance compared to prior methods. This work establishes a novel paradigm for medical image synthesis and in silico experimentation.

Contrastive learning in self-supervised 3D voxel shape representation learning often suffers from latent representation collapse. Method: We propose a generative-contrastive joint learning framework featuring a dual-branch encoder—processing voxels and multi-view images separately—coupled with a shared decoder and a switching training mechanism. To mitigate representation degradation, we introduce randomized stop-gradient operations. The framework jointly optimizes cross-modal contrastive loss and voxel reconstruction loss to achieve implicit feature alignment. Results: Experiments demonstrate substantial improvements over pure contrastive baselines on downstream classification and reconstruction tasks. Our approach effectively alleviates representation collapse, enhancing both discriminability and geometric fidelity of multimodal representations. It establishes a scalable, bi-modal collaborative paradigm for 3D self-supervised learning.

This study addresses the decoding distortion caused by latent space geometric anisotropy in 3D medical image synthesis by proposing a Latent Structure Flow (LSF) method. LSF decouples latent states into structural and residual components to model global and local variations separately. By revealing that structural variations are governed by low-rank subspaces, we design an efficient architecture that modifies only the generator while freezing the pretrained encoder-decoder to reduce optimization complexity. Experimental results demonstrate that LSF consistently outperforms existing baselines across multimodal medical image synthesis and tumor inpainting tasks, achieving high-quality and structurally consistent 3D medical image generation.

研究对比了解剖特征与学习特征在脑MRI分析中的效果，提出结合解剖信息预训练的新方法，提升了生物年龄估计的准确性。

Existing cross-modal medical image translation methods typically rely on 2D slices or 3D local patches and require separate model training for each task, resulting in limited generalization. This work proposes a whole-volume representation learning framework based on a 3D variational autoencoder, formulating cross-modal translation as a conditional flow-matching problem in latent space. By integrating resolution-aware sampling with multi-task joint training, the approach enables a single model to support diverse modality conversions. It achieves, for the first time, unified whole-volume, multi-task translation with less than 0.15 SSIM performance drop on unseen anatomical regions. The method further supports zero-shot anatomical generalization and unsupervised cross-dataset compositional translation, matching the performance of task-specific models.

Standard CNNs for whole-heart multi-chamber CT segmentation often lack explicit anatomical constraints, compromising clinical reliability. This work proposes a lightweight approach that explicitly incorporates statistical shape priors through a shape-aware loss and a 3D U-Net variant guided by spatial label distribution heatmaps. The method is systematically evaluated on the MM-WHS CT and WHS++ datasets. Results reveal that, despite modern architectures implicitly learning substantial anatomical regularities from data, explicitly integrating handcrafted shape priors yields only marginal and inconsistent performance gains—and frequently leads to degradation. These findings underscore the limited added value of manually designed anatomical priors when deployed within highly data-driven deep learning models.

研究通过引入解剖信息神经网络解决深度学习模型在稀缺数据下解剖学预测不准确的问题，使用SE(3)公式化方法处理导丝引起的主动脉髂动脉变形。