Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reconsider the Template Mesh in Deep Learning-based Mesh Reconstruction

Fengting Zhang, Boxu Liang, Qinghao Liu, Min Liu, Xiang Chen, Yaonan Wang

TL;DR

The paper addresses the limitation of fixed templates in deep learning-based mesh reconstruction by introducing an adaptive Template-based Mesh Reconstruction Network (ATMRN). ATMRN first generates an adaptive template T_a = T_s + T_d from image features via a U-Net-like encoder and a GCN mesh decoder, then deforms this template with volume-to-point-cloud mapping and staged GCN refinement to the target cortex mesh, optimized by a multi-term mesh loss L_mesh = λ1 L_CD + λ2 L_Laplacian + λ3 L_normal + λ4 L_edge. Key contributions include (1) introducing adaptive templates within a DL mesh reconstruction framework, (2) achieving state-of-the-art cortical mesh reconstruction on the OASIS dataset with average surface distance around 0.267 mm, and (3) providing a systematic analysis of template choices to guide future work. The approach demonstrates that subject-specific adaptive templates improve reconstruction fidelity and generalize to other imaging modalities and anatomical structures, with practical impact for neuroimaging and surgical planning.

Abstract

Mesh reconstruction is a cornerstone process across various applications, including in-silico trials, digital twins, surgical planning, and navigation. Recent advancements in deep learning have notably enhanced mesh reconstruction speeds. Yet, traditional methods predominantly rely on deforming a standardised template mesh for individual subjects, which overlooks the unique anatomical variations between them, and may compromise the fidelity of the reconstructions. In this paper, we propose an adaptive-template-based mesh reconstruction network (ATMRN), which generates adaptive templates from the given images for the subsequent deformation, moving beyond the constraints of a singular, fixed template. Our approach, validated on cortical magnetic resonance (MR) images from the OASIS dataset, sets a new benchmark in voxel-to-cortex mesh reconstruction, achieving an average symmetric surface distance of 0.267mm across four cortical structures. Our proposed method is generic and can be easily transferred to other image modalities and anatomical structures.

Reconsider the Template Mesh in Deep Learning-based Mesh Reconstruction

TL;DR

The paper addresses the limitation of fixed templates in deep learning-based mesh reconstruction by introducing an adaptive Template-based Mesh Reconstruction Network (ATMRN). ATMRN first generates an adaptive template T_a = T_s + T_d from image features via a U-Net-like encoder and a GCN mesh decoder, then deforms this template with volume-to-point-cloud mapping and staged GCN refinement to the target cortex mesh, optimized by a multi-term mesh loss L_mesh = λ1 L_CD + λ2 L_Laplacian + λ3 L_normal + λ4 L_edge. Key contributions include (1) introducing adaptive templates within a DL mesh reconstruction framework, (2) achieving state-of-the-art cortical mesh reconstruction on the OASIS dataset with average surface distance around 0.267 mm, and (3) providing a systematic analysis of template choices to guide future work. The approach demonstrates that subject-specific adaptive templates improve reconstruction fidelity and generalize to other imaging modalities and anatomical structures, with practical impact for neuroimaging and surgical planning.

Abstract

Mesh reconstruction is a cornerstone process across various applications, including in-silico trials, digital twins, surgical planning, and navigation. Recent advancements in deep learning have notably enhanced mesh reconstruction speeds. Yet, traditional methods predominantly rely on deforming a standardised template mesh for individual subjects, which overlooks the unique anatomical variations between them, and may compromise the fidelity of the reconstructions. In this paper, we propose an adaptive-template-based mesh reconstruction network (ATMRN), which generates adaptive templates from the given images for the subsequent deformation, moving beyond the constraints of a singular, fixed template. Our approach, validated on cortical magnetic resonance (MR) images from the OASIS dataset, sets a new benchmark in voxel-to-cortex mesh reconstruction, achieving an average symmetric surface distance of 0.267mm across four cortical structures. Our proposed method is generic and can be easily transferred to other image modalities and anatomical structures.

Paper Structure

This paper contains 8 sections, 3 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Network Architecture
  4. Loss Function.
  5. Experiments and Results
  6. Experiments Settings
  7. Results and Analysis
  8. Conclusion

Figures (2)

  • Figure 1: Schema of our proposed ATMRN. Our proposed ATMRN comprises four blocks, the feature extraction, GCN mesh decoder, volume-to-PC mapping and GCN mesh deformation blocks.
  • Figure 2: Qualitative results between ATMRN and SOTA methods. The ground-truth meshes are on the left (from the top to bottom are the surface meshes of left WM, right WM, left pial and right pial), with the corresponding predicted meshes on the right. The colour bar denotes the distance between the predicted and ground-truth meshes.