Low-Dose CT With a Residual Encoder-Decoder Convolutional Neural Network

Hu Chen, Yi Zhang, Mannudeep K. Kalra, Feng Lin, Yang Chen, Peixi Liao, Jiliu Zhou, Ge WangIEEE TMI 2017 · OpenReview

Abstract

Given the potential risk of X-ray radiation to the patient, low-dose CT has attracted a considerable interest in the medical imaging field. Currently, the main stream low-dose CT methods include vendor-specific sinogram domain filtration and iterative reconstruction algorithms, but they need to access raw data, whose formats are not transparent to most users. Due to the difficulty of modeling the statistical characteristics in the image domain, the existing methods for directly processing reconstructed images cannot eliminate image noise very well while keeping structural details. Inspired by the idea of deep learning, here we combine the autoencoder, deconvolution network, and shortcut connections into the residual encoder-decoder convolutional neural network (RED-CNN) for low-dose CT imaging. After patch-based training, the proposed RED-CNN achieves a competitive performance relative to the-state-of-art methods in both simulated and clinical cases. Especially, our method has been favorably evaluated in terms of noise suppression, structural preservation, and lesion detection.

Introduction

Computed Tomography (CT) is a vital diagnostic tool, but the X-ray radiation involved poses health risks. To follow the "ALARA" (As Low As Reasonably Achievable) principle, clinicians try to use low-dose scans. However, reducing the X-ray flux introduces significant noise and "streaking" artifacts that can obscure small lesions or blood vessels.

Existing solutions generally fall into two camps: sinogram/iterative methods and image-domain processing. Sinogram methods work on raw scanner data but are often slow and vendor-specific. Image-domain processing works on final reconstructed images (like DICOMs) but traditional filters often blur the image, losing sharp edges needed for diagnosis. The Residual Encoder-Decoder Convolutional Neural Network (RED-CNN) is designed to denoise low-dose CT (LDCT) scans directly in the image domain, suppressing noise and artifacts while preserving critical diagnostic details.

Methods

The proposed RED-CNN architecture features several key innovations designed for high-fidelity medical imaging.

Architecture of the Residual Encoder-Decoder Convolutional Neural Network (RED-CNN).

Architecture

  • Symmetric Encoder-Decoder: The network consists of 10 layers—5 convolutional layers (the encoder) to extract features and 5 deconvolutional layers (the decoder) to reconstruct the image.
  • Residual Shortcuts: Inspired by ResNet, shortcuts connect a convolutional layer directly to its corresponding deconvolutional layer. This allows the network to pass high-frequency structural information directly to the output, making it easier to learn the "residual" (the noise) rather than having to relearn the entire image from scratch.
  • No Pooling Layers: Unlike standard computer vision models, RED-CNN discards pooling or down-sampling layers. In medical imaging, every pixel matters; pooling would discard tiny details like calcifications.

Training and Datasets

The model is trained on small $55 \times 55$ image patches to learn local textures and increase the effective dataset size. Two primary datasets were utilized: the Simulated Data (NBIA), consisting of 7,015 images with mathematically added Poisson noise, and the Clinical Data (Mayo Clinic), containing 2,378 real-world pairs of full-dose and quarter-dose CT images.

Results and Discussion

RED-CNN was compared against five state-of-the-art methods, including TV-POCS (iterative) and BM3D (traditional filtering).

Quantitative and Qualitative Performance

RED-CNN achieved the highest Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) across all tests. In the Mayo Clinic dataset, it reached a PSNR of ~43.79 dB, significantly higher than the ~36.40 dB of the raw low-dose images. Visually, RED-CNN was the only method that successfully preserved the subtle margins of the adrenal glands and small blood vessels in the lungs while completely removing bone-induced streaks.

Efficiency and Robustness

A major clinical advantage is speed. While iterative reconstruction (ASD-POCS) took over 21 seconds per slice, RED-CNN processed a slice in just 3.68 seconds on a standard GPU. Furthermore, by training on a mix of different noise levels (RED-CNN+), the model became robust enough to handle scans from various hardware or dose settings without retraining.

Conclusion

RED-CNN demonstrates that a deep learning approach using a symmetric encoder-decoder structure with residual connections can effectively bridge the gap between low-dose safety and high-quality diagnostic imaging. While currently a 2D model, future work will explore 3D convolutions to leverage information from adjacent slices and extend the framework to dynamic and spectral CT imaging.

Summary updated

Abstract

Given the potential risk of X-ray radiation to the patient, low-dose CT has attracted a considerable interest in the medical imaging field. Currently, the main stream low-dose CT methods include vendor-specific sinogram domain filtration and iterative reconstruction algorithms, but they need to access raw data, whose formats are not transparent to most users. Due to the difficulty of modeling the statistical characteristics in the image domain, the existing methods for directly processing reconstructed images cannot eliminate image noise very well while keeping structural details. Inspired by the idea of deep learning, here we combine the autoencoder, deconvolution network, and shortcut connections into the residual encoder-decoder convolutional neural network (RED-CNN) for low-dose CT imaging. After patch-based training, the proposed RED-CNN achieves a competitive performance relative to the-state-of-art methods in both simulated and clinical cases. Especially, our method has been favorably evaluated in terms of noise suppression, structural preservation, and lesion detection.
BibTeX
@article{chen2017low,
  author = {Hu Chen and Yi Zhang and Mannudeep K. Kalra and Feng Lin and Yang Chen and Peixi Liao and Jiliu Zhou and Ge Wang},
  title = {{Low-Dose CT With a Residual Encoder-Decoder Convolutional Neural Network}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2017},
  url = {https://openreview.net/forum?id=vZx1pEngyaY},
}
Figure