Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Deep MR to CT Synthesis using Unpaired Data (1708.01155v1)

Published 3 Aug 2017 in cs.CV

Abstract: MR-only radiotherapy treatment planning requires accurate MR-to-CT synthesis. Current deep learning methods for MR-to-CT synthesis depend on pairwise aligned MR and CT training images of the same patient. However, misalignment between paired images could lead to errors in synthesized CT images. To overcome this, we propose to train a generative adversarial network (GAN) with unpaired MR and CT images. A GAN consisting of two synthesis convolutional neural networks (CNNs) and two discriminator CNNs was trained with cycle consistency to transform 2D brain MR image slices into 2D brain CT image slices and vice versa. Brain MR and CT images of 24 patients were analyzed. A quantitative evaluation showed that the model was able to synthesize CT images that closely approximate reference CT images, and was able to outperform a GAN model trained with paired MR and CT images.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (6)
  1. Jelmer M. Wolterink (56 papers)
  2. Anna M. Dinkla (2 papers)
  3. Mark H. F. Savenije (2 papers)
  4. Peter R. Seevinck (4 papers)
  5. Cornelis A. T. van den Berg (23 papers)
  6. Ivana Isgum (64 papers)
Citations (580)

Summary

Deep MR to CT Synthesis using Unpaired Data

The paper "Deep MR to CT Synthesis using Unpaired Data" by Wolterink et al. presents a novel approach to synthesizing CT images from MR images using unpaired datasets. The paper addresses the limitations of previous methodologies that rely heavily on aligned MR and CT image pairs, often suffering from misalignment errors that can degrade the quality of synthesized images.

Methodological Overview

Utilizing a CycleGAN framework, the authors propose a solution that bypasses the need for paired training data. The CycleGAN model consists of two synthesis CNNs and two discriminator CNNs, trained with cycle consistency to achieve bidirectional transformation between 2D brain MR and CT image slices. This model introduces a comprehensive cycle consistency loss, which ensures that a generated image can be effectively transformed back to its original domain, thereby maintaining the integrity of the generated images.

Data Utilization

The research employed MR and CT images from 24 patients, specifically chosen for radiotherapy treatment planning. To facilitate voxel-wise comparison, MR and CT images from the same patients were aligned post acquisition using rigid registration. By focusing on unpaired images, the paper opens avenues for using datasets that are typically challenging to acquire in aligned pairs, thereby enhancing the practical applicability of the method.

Experimental Findings

The paper's empirical results indicate a notable improvement with this unpaired training approach. The synthesized CT images achieved a mean absolute error (MAE) of 73.7 HU and a peak-signal-to-noise ratio (PSNR) of 32.3 compared to reference CT images. Notably, these results outperform those obtained from a GAN model trained on paired data, where artifacts and blurring were more pronounced. The findings suggest potential advances in medical imaging, particularly in scenarios where training data is naturally unaligned or scarce.

Implications and Future Directions

The unpaired model significantly reduces requirements for spatial alignment in training data, suggesting a practical advantage in clinical environments. This flexibility is particularly beneficial in domains with diverse imaging modalities, such as MR imaging at varying field strengths or CT imaging at different dose levels. Furthermore, the integration of higher-dimensional (3D) MR and CT data, as proposed for future work, could further refine model performance by leveraging spatial continuity and depth information absent in 2D training.

In summary, this approach offers a feasible alternative to traditional paired-data models, broadening access and reducing dependency on rigidly aligned datasets, ultimately contributing to improved outcomes in MR-only radiotherapy treatment planning and beyond.