PaperlayerView exact revision on arXiv

Source-linked AI summary

TransCT: Dual-path Transformer for Low Dose Computed Tomography

Zhicheng Zhang, Lequan Yu, Xiaokun Liang, Wei Zhao, Lei Xing

arXiv:2103.00634v4eess.IVphysics.med-ph

TL;DR

Reducing X-ray exposure lowers radiation dose but degrades CT image quality. TransCT splits noisy LDCT into HF and LF components, uses LF textures to refine HF features, and reconstructs images piecewise. On the Mayo dataset, the authors report improved image-quality metrics over LDCT and superior performance and robustness relative to other methods.

  • Problem

    Lowering X-ray exposure reduces radiation dose but degrades CT image quality, motivating methods that preserve image quality at lower dose.

    Source: §1 · p. 1

  • Method

    TransCT decomposes LDCT into HF and LF parts, extracts LF content and texture features and HF embeddings, then uses a modified transformer and piecewise reconstruction.

    Source: §1 · p. 2

  • Results

    Against LDCT, TransCT decreases RMSE by 40.5%, improves SSIM by 12.3%, and VIF by 93.7%.

    Source: Quantitative Analysis · p. 6

  • Takeaways & Limitations

    On the Mayo dataset, the authors report that TransCT outperforms other methods and has superior robustness.

    Source: Quantitative Analysis · p. 7

Abstract

from arXiv · show

Low dose computed tomography (LDCT) has attracted more and more attention in routine clinical diagnosis assessment, therapy planning, etc., which can reduce the dose of X-ray radiation to patients. However, the noise caused by low X-ray exposure degrades the CT image quality and then affects clinical diagnosis accuracy. In this paper, we train a transformer-based neural network to enhance the final CT image quality. To be specific, we first decompose the noisy LDCT image into two parts: high-frequency (HF) and low-frequency (LF) compositions. Then, we extract content features (X_{L_c}) and latent texture features (X_{L_t}) from the LF part, as well as HF embeddings (X_{H_f}) from the HF part. Further, we feed X_{L_t} and X_{H_f} into a modified transformer with three encoders and decoders to obtain well-refined HF texture features. After that, we combine these well-refined HF texture features with the pre-extracted X_{L_c} to encourage the restoration of high-quality LDCT images with the assistance of piecewise reconstruction. Extensive experiments on Mayo LDCT dataset show that our method produces superior results and outperforms other methods.

1 Introduction

Reducing X-ray exposure is a practical way to lower radiation dose, but lower exposure degrades CT images. TransCT uses LF latent textures to refine HF features and reconstruct LDCT images.

1 Introduction

  • Lowering single X-ray exposure is described as the most convenient and affordable dose-reduction strategy, though it degrades CT image quality.The paper frames preserving image quality while reducing radiation dose as the research goal.Source: §1 · p. 1
  • Prior LDCT approaches include model-based methods limited by model-description accuracy and data-driven reconstruction methods, including CNN and iterative-learning approaches.The paper also notes AUTOMAP’s first fully connected layer has high memory requirements for large-scale CT reconstruction.Source: §1 · pp. 1–2
  • CNN-based LDCT methods rely on cascaded convolutions with limited receptive fields, hindering modeling of long-range dependencies and structural information.Transformers can model interactions between distant image regions through attention.Source: §1 · p. 2
  • TransCT decomposes noisy LDCT into HF and LF parts, using LF content and latent textures alongside HF embeddings in a transformer-based restoration pipeline.The refined HF features are combined with LF content for piecewise reconstruction.Source: §1 · pp. 2–3

2 Method

TransCT separates noisy LDCT into frequency components, extracts LF content and texture features, and uses a modified transformer to refine HF features. It combines the refined features with LF content through staged reconstruction and trains against NDCT using MSE.

2 Method

  • A Gaussian filter with standard deviation 1.5 decomposes the noisy LDCT image into HF and LF components.The method uses LF latent textures to assist noise removal in the HF part.Source: §2 · p. 3

2.1 TransCT

  • Two shallow CNN paths extract content features and latent texture features from the LF image for subsequent reconstruction and transformer processing.The architecture also extracts HF embeddings from the HF component.Source: §2 · p. 3, Sequence · p. 4

Sequence

  • The LF texture features and HF embeddings are reshaped into separate sequences as transformer inputs.The sequences provide token representations for the LF and HF components.Source: Sequence · p. 4

Transformer

  • The modified transformer has three encoders and three decoders; decoder attention uses the final LF texture sequence as its key and value.The encoder processes LF tokens, while the HF sequence is passed through decoder attention.Source: Transformer · p. 4

Piecewise Reconstruction

  • The transformer output is combined with LF content features and reconstructed in stages using residual blocks and sub-pixel upsampling.The first stage combines output Y with XLc2; the next combines the upsampled features with XLc1.Source: Piecewise Reconstruction · pp. 4–5

2.2 Loss Function

  • The model is trained with MSE between its output and the normal-dose CT image.The passage identifies the normal-dose image as IND, the low-dose input as ILD, and θ as the network parameters.Source: §2.2 · p. 5

2.3 Implementation

  • Training used Adam for 300 epochs with batch size 8 and a learning rate reduced from 0.0001 to 0.00001 after 180 epochs.The implementation was based on TensorFlow and used Xavier initialization.Source: §2.3 · p. 5

3 Experiments

Experiments used Mayo data and compared TransCT with established LDCT baselines. The authors report quantitative and visual gains, then examine piecewise reconstruction, model size, and module choices.

Datasets

  • The dataset contains normal-dose and simulated quarter-dose abdominal CT from 10 patients, split into 7 training, 1 validation, and 2 testing cases.The comparison baselines were NLM, RED-CNN, and MAP-NN; the latter two were retrained on the same dataset.Source: Datasets · p. 5, Comparison with other methods · p. 5

Comparison with other methods

Quantitative Analysis

  • With tube currents of 80mA for NDCT and 20mA for LDCT, the authors report that TransCT outperforms other methods with superior robustness.Source: Quantitative Analysis · p. 7

On the Influence of Piecewise Reconstruction

On the Influence of Model Size

  • Replacing the transformer module with convolutional/residual blocks or removing the HF path reduced validation performance relative to TransCT with both modules.The ablations are reported in Table 2.Source: On the Influence of Model Size · p. 8

4 Conclusion

TransCT applies a transformer to LDCT by exploiting long-range pixel dependencies and using low-frequency texture features to restore high-frequency features. The authors propose future self-supervised learning to reduce training-data needs.

4 Conclusion

  • TransCT is presented as the first transformer-based LDCT network, designed to explore long-range dependencies between image pixels.Source: §4 · p. 8
  • The method uses low-frequency latent texture features to restore noisy high-frequency features, then combines low-frequency content and high-frequency features in piecewise reconstruction.The authors identify self-supervised learning as future work to lower training-data needs.Source: §4 · p. 9

Related papers

Loading 2103.00634v4…