Abstract:

3D modeling of the anatomical structures based on 3D scan information is essential in the production of customized medical instruments. Raw point cloud data is however, in most cases quite noisy, incomplete and is sampled quite irregularly making it difficult to reconstruct with much reliability. This paper presents a comparative analysis of advanced deep learning approaches for 3D reconstruction, including a pointbased model, PointNet++, and a Transformer-based implicit surface reconstruction model, as well as a proposed hybrid framework that integrates both paradigms. The performance of these approaches is evaluated using standard geometric metrics,
including Chamfer Distance, Hausdorff Distance, and Normal Consistency, in addition to classification-based measures such as Precision, Recall, and F-score. Experimental results demonstrate that while PointNet++ effectively preserves local geometric details and achieves high precision, the Transformer-based model provides improved global context modeling and surface smoothness, and the proposed hybrid model achieves the best overall performance. Specifically, it provides a balanced trade-off between geometric accuracy, surface continuity, and structural consistency, outperforming individual models across multiple evaluation criteria. These findings highlight the advantages of integrating complementary 3D representation paradigms and
provide practical insights for selecting appropriate reconstruction techniques in the automated and scalable design of personalized anatomical models, reducing the need for manual intervention in medical applications. Index Terms—3D reconstruction