Low-Dose CT via Deep Neural Network

Hu Chen, Yi Zhang, Weihua Zhang, Peixi Liao, Ke Li, Jiliu Zhou, Ge WangarXiv 2016 · DOI · arXiv

Abstract

In order to reduce the potential radiation risk, low-dose CT has attracted more and more attention. However, simply lowering the radiation dose will significantly degrade the imaging quality. In this paper, we propose a noise reduction method for low-dose CT via deep learning without accessing the original projection data. An architecture of deep convolutional neural network was considered to map the low-dose CT images into its corresponding normal-dose CT images patch by patch. Qualitative and quantitative evaluations demonstrate a state-the-art performance of the proposed method.

Introduction

X-ray Computed Tomography (CT) is an indispensable diagnostic tool in modern medicine. However, the ionizing radiation associated with CT scans poses potential health risks, including an increased lifetime risk of cancer. While reducing the radiation dose is a clinical priority, it introduces significant "quantum noise," resulting in streaks and artifacts that can compromise diagnostic accuracy.

Current noise reduction strategies typically fall into three categories: sinogram filtering, iterative reconstruction, and post-reconstruction image processing. Sinogram and iterative methods often require access to raw projection data, which is frequently restricted by proprietary vendor encryption. Furthermore, iterative methods are computationally demanding and slow. This paper proposes a deep learning-based image processing approach that operates directly on reconstructed images, bypassing the need for raw data and offering high computational efficiency.

Methodology

The proposed method utilizes a three-layer Convolutional Neural Network (CNN) architecture designed to learn the mapping from noisy low-dose images to clean normal-dose counterparts.

  1. Patch Encoding: The first layer extracts small patches (33x33 pixels) from the input image and maps them into a high-dimensional feature space. This stage is conceptually similar to sparse representation, where signals are decomposed into a set of basis functions.
  2. Non-linear Mapping: The second layer performs a non-linear transformation, mapping the noisy feature vectors into a "clean" feature space.
  3. Reconstruction: The final layer aggregates these processed features to reconstruct the denoised image.

The network is trained using Mean Squared Error (MSE) as the loss function. A key advantage of this approach is that it does not require an explicit mathematical model of CT noise; instead, it learns the complex, non-Gaussian noise characteristics directly from the data.

Datasets and Experimental Setup

The study leveraged a combination of simulated and real-world data to train and validate the model. The primary training set consisted of 7,015 normal-dose CT images from 165 patients, obtained from the National Cancer Imaging Archive (NCIA).

Representative CT slices from the NCIA dataset used in the study.

To generate training pairs, low-dose images were simulated by applying fan-beam geometry and adding Poisson noise to the projection domain of the clean images. Additionally, the model's generalizability was tested using real-world low-dose data from a sheep lung perfusion study conducted on a Siemens Somatom scanner.

Key Results

The CNN's performance was benchmarked against several established methods: ASD-POCS (iterative), K-SVD (dictionary learning), and BM3D (traditional denoising).

Quantitative Evaluation

The proposed CNN consistently outperformed competing methods across standard metrics. On a representative test set, the CNN achieved a Peak Signal-to-Noise Ratio (PSNR) of 42.15 dB, significantly higher than the low-dose baseline (36.39 dB) and BM3D (41.53 dB). It also recorded the lowest Root Mean Square Error (RMSE).

Qualitative Evaluation

Visual assessments by experienced radiologists confirmed the quantitative findings. While iterative methods like ASD-POCS often produced "blocky" artifacts and BM3D struggled with streaks near high-contrast bone structures, the CNN effectively suppressed noise while preserving sharp boundaries and fine structural details.

Comparison of denoising performance across different methods, with zoomed-in regions showing detail preservation.

In the sheep lung study, the CNN was uniquely successful in eliminating streak artifacts without blurring the delicate lung tissue architecture.

Discussion and Conclusion

The proposed deep learning framework offers a robust and efficient solution for low-dose CT denoising. By processing images patch-by-patch, the network captures local structural information and achieves high throughput, with an inference time of approximately 2 seconds per image on a standard GPU.

Despite its success, the authors noted that BM3D occasionally showed slightly better Structural Similarity (SSIM) in soft-tissue abdominal scans, suggesting that further optimization for low-contrast regions may be beneficial. Future work will explore the application of this architecture to other CT challenges, such as metal artifact reduction and interior CT reconstruction.

Summary updated

Abstract

In order to reduce the potential radiation risk, low-dose CT has attracted more and more attention. However, simply lowering the radiation dose will significantly degrade the imaging quality. In this paper, we propose a noise reduction method for low-dose CT via deep learning without accessing the original projection data. An architecture of deep convolutional neural network was considered to map the low-dose CT images into its corresponding normal-dose CT images patch by patch. Qualitative and quantitative evaluations demonstrate a state-the-art performance of the proposed method.
BibTeX
@misc{chen2016low,
  author = {Hu Chen and Yi Zhang and Weihua Zhang and Peixi Liao and Ke Li and Jiliu Zhou and Ge Wang},
  title = {{Low-Dose CT via Deep Neural Network}},
  year = {2016},
  eprint = {1609.08508},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1609.08508},
}
Figure