Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Cross-Modality Synthesis: Subvolume Merging for MRI-to-CT Conversion

Fuxin Fan, Jingna Qiu, Yixing Huang, Andreas Maier

TL;DR

This work tackles MRI-to-CT synthesis for radiotherapy planning by combining a SwinUNETR-based network with a novel 3D subvolume merging strategy during inference to reduce stitching artifacts and improve attenuation accuracy. The method uses $32\times96\times96$ subvolumes and a patch size of $2\times2\times2$ within a four-stage Swin Transformer–enabled encoder–decoder, with L1 loss and Adam optimization. Empirical results show the subvolume merging approach lowers MAE from $52.65\ \mathrm{HU}$ to $47.75\ \mathrm{HU}$ and that a gamma-weighted overlap with $\gamma=0.9$ yields optimal performance; an overlap of $50\%$–$70\%$ balances image quality and computational efficiency. The findings support broader applicability to other regression tasks requiring subvolume processing in medical imaging.

Abstract

Providing more precise tissue attenuation information, synthetic computed tomography (sCT) generated from magnetic resonance imaging (MRI) contributes to improved radiation therapy treatment planning. In our study, we employ the advanced SwinUNETR framework for synthesizing CT from MRI images. Additionally, we introduce a three-dimensional subvolume merging technique in the prediction process. By selecting an optimal overlap percentage for adjacent subvolumes, stitching artifacts are effectively mitigated, leading to a decrease in the mean absolute error (MAE) between sCT and the labels from 52.65 HU to 47.75 HU. Furthermore, implementing a weight function with a gamma value of 0.9 results in the lowest MAE within the same overlap area. By setting the overlap percentage between 50% and 70%, we achieve a balance between image quality and computational efficiency.

Enhancing Cross-Modality Synthesis: Subvolume Merging for MRI-to-CT Conversion

TL;DR

This work tackles MRI-to-CT synthesis for radiotherapy planning by combining a SwinUNETR-based network with a novel 3D subvolume merging strategy during inference to reduce stitching artifacts and improve attenuation accuracy. The method uses subvolumes and a patch size of within a four-stage Swin Transformer–enabled encoder–decoder, with L1 loss and Adam optimization. Empirical results show the subvolume merging approach lowers MAE from to and that a gamma-weighted overlap with yields optimal performance; an overlap of – balances image quality and computational efficiency. The findings support broader applicability to other regression tasks requiring subvolume processing in medical imaging.

Abstract

Providing more precise tissue attenuation information, synthetic computed tomography (sCT) generated from magnetic resonance imaging (MRI) contributes to improved radiation therapy treatment planning. In our study, we employ the advanced SwinUNETR framework for synthesizing CT from MRI images. Additionally, we introduce a three-dimensional subvolume merging technique in the prediction process. By selecting an optimal overlap percentage for adjacent subvolumes, stitching artifacts are effectively mitigated, leading to a decrease in the mean absolute error (MAE) between sCT and the labels from 52.65 HU to 47.75 HU. Furthermore, implementing a weight function with a gamma value of 0.9 results in the lowest MAE within the same overlap area. By setting the overlap percentage between 50% and 70%, we achieve a balance between image quality and computational efficiency.
Paper Structure (7 sections, 2 equations, 5 figures)

This paper contains 7 sections, 2 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Methods and Materials
  3. Neural network
  4. Subvolume merging
  5. Data set preparation
  6. Results and discussion
  7. Conclusion

Figures (5)

  • Figure 1: The network structure of SwinUNETR.
  • Figure 2: 3D subvolume merging during the inference process.
  • Figure 3: The result of sCT before and after subvolume merging. The right two columns show the enlarged regions from the 3rd and 4th columns. Such regions are highlighted by blue arrows in the third column. The intensity window for CT images is [-500, 500] HU.
  • Figure 4: MAE with respect to Gamma values.
  • Figure 5: Ablation study on overlap percentage.