A deep convolutional neural network using directional wavelets for low-dose X-ray CT reconstruction

Abstract: Due to the potential risk of inducing cancers, radiation dose of X-ray CT should be reduced for routine patient scanning. However, in low-dose X-ray CT, severe artifacts usually occur due to photon starvation, beamhardening, etc, which decrease the reliability of diagnosis. Thus, high quality reconstruction from low-dose X-ray CT data has become one of the important research topics in CT community. Conventional model-based denoising approaches are, however, computationally very expensive, and image domain denoising approaches hardly deal with CT specific noise patterns. To address these issues, we propose an algorithm using a deep convolutional neural network (CNN), which is applied to wavelet transform coefficients of low-dose CT images. Specifically, by using a directional wavelet transform for extracting directional component of artifacts and exploiting the intra- and inter-band correlations, our deep network can effectively suppress CT specific noises. Moreover, our CNN is designed to have various types of residual learning architecture for faster network training and better denoising. Experimental results confirm that the proposed algorithm effectively removes complex noise patterns of CT images, originated from the reduced X-ray dose. In addition, we show that wavelet domain CNN is efficient in removing the noises from low-dose CT compared to an image domain CNN. Our results were rigorously evaluated by several radiologists and won the second place award in 2016 AAPM Low-Dose CT Grand Challenge. To the best of our knowledge, this work is the first deep learning architecture for low-dose CT reconstruction that has been rigorously evaluated and proven for its efficacy.

• The paper introduces a deep CNN that combines directional wavelet transforms with residual learning to effectively suppress CT-specific noise.
• The network features 24 convolution layers and bypass connections, validated with rigorous testing against the 2016 AAPM Low-Dose CT Grand Challenge dataset.
• It offers a computationally efficient alternative to traditional MBIR methods, preserving fine image details and reducing artifacts in low-dose CT images.

Deep CNNs for Low-Dose CT Reconstruction

This paper introduces a novel deep convolutional neural network (CNN) architecture for low-dose X-ray computed tomography (CT) reconstruction, leveraging directional wavelets to address CT-specific noise patterns. The algorithm combines a deep CNN with a directional wavelet transform to effectively suppress noise while preserving image details, offering a computationally efficient alternative to traditional model-based iterative reconstruction (MBIR) methods. The method's efficacy was validated through rigorous evaluation and demonstrated promising results in the 2016 AAPM Low-Dose CT Grand Challenge.

Background and Motivation

The primary motivation behind this research is the need to minimize radiation exposure in routine CT scans due to the potential risk of radiation-induced cancer. Low-dose CT imaging, however, introduces significant artifacts such as photon starvation and beam hardening, which can compromise diagnostic reliability. Conventional image-domain denoising techniques often struggle with CT-specific noise patterns, while MBIR methods are computationally intensive. The proposed algorithm aims to overcome these limitations by using a deep-learning approach that exploits the characteristics of directional wavelets and the capabilities of CNNs.

Figure 1: Various noise patterns in low-dose CT images: (a) Gaussian noise, and (b) streaking artifacts.

The paper acknowledges that low-dose CT images are characterized by complex noise patterns, including Gaussian noise and streaking artifacts (Figure 1). Traditional denoising methods, such as total variation minimization and wavelet shrinkage, have limitations in addressing these specific artifacts. MBIR algorithms, while more tailored to CT imaging, involve computationally expensive iterative projection/backprojection steps.

Methodology: Wavelet-Domain Deep CNN

The proposed algorithm leverages a deep CNN applied to the wavelet transform coefficients of low-dose CT images. A directional wavelet transform extracts the directional components of artifacts, and the deep network exploits intra- and inter-band correlations to suppress CT-specific noise effectively. The CNN is designed with a residual learning architecture for faster network training and enhanced performance. The network architecture incorporates several key components, including convolution layers, batch normalization, and rectified linear units (ReLU), with various bypass connections to facilitate residual learning.

Figure 2: Non-subsampled contourlet transform: (a) Scheme of contourlet transform. First, the image is split into high-pass and low-pass subbands. Then, non-subsampled directional filter banks divide the high-pass subband into directional subbands. This process is repeated in the low-pass subband. (b) Examples of the non-subsampled contourlet transform of a low-dose CT image. There are four levels, one each with eight, four, two, and one directional subbands.

The contourlet transform (Figure 2) is used to decompose the input noisy image into four decomposition levels, generating a total of 15 channels. The noisy wavelet coefficients are processed in a patch-by-patch manner using a convolution operator, with each patch consisting of 55×55 square regions from the 15 channels. The network employs residual learning, where low-frequency wavelet coefficients are bypassed and later added to the denoised wavelet coefficients. Internal bypass connections further aid in training the deep network and improving denoising performance.

Figure 3: Proposed deep convolutional neural network architecture for wavelet domain de-noising

The proposed network contains 24 convolution layers, each followed by a batch normalization layer and a ReLU layer (Figure 3). 128 sets of 3×3×15 convolution filters are used on the first layer, followed by 128 sets of 3×3×128 convolution filters in subsequent layers. The network consists of six modules, each containing a bypass connection and three convolution layers. A channel concatenation layer stacks the inputs of each module in the channel dimension, enabling faster end-to-end training.

Experimental Evaluation and Results

The algorithm's performance was rigorously evaluated using the 2016 Low-Dose CT Grand Challenge dataset. The experimental results confirm that the proposed algorithm effectively removes complex noise patterns from CT images derived from reduced X-ray doses. The wavelet-domain CNN demonstrates efficiency in removing noise from low-dose CT images compared to existing approaches. The results were evaluated by radiologists at the Mayo Clinic, leading to a second-place win in the challenge.

Figure 4: Reconstruction results from the test data `L031': (a) quarter-dose image, and the results by (b) the proposed network trained with 1mm slices followed by 3mm averaging, and (c) the proposed network trained with 3mm slices. The second column shows enlarged images from the yellow boxes. Yellow arrows indicate the image details. The intensity range was set to (-160,240) [HU].

The paper presents visual comparisons of reconstructed images from both training and test datasets, demonstrating the algorithm's ability to preserve fine image details and suppress streaking artifacts (Figure 4). Quantitative assessments using peak signal-to-noise ratio (PSNR) and normalized root mean square error (NRMSE) values further validate the effectiveness of the proposed network. The authors also analyze the role of the wavelet transform and residual learning techniques through comparative experiments with baseline CNN architectures.

Discussion and Implications

The paper highlights the advantage of the proposed network over MBIR approaches in fully utilizing large training datasets. While MBIR methods typically train a single regularization parameter, the proposed network can train a large number of neural network parameters more accurately with more data. This capability to learn organ, protocol, and hardware-dependent noise from larger datasets offers a significant advantage.

The authors acknowledge that the current network is adapted to a specific quarter-dose level, and additional training with data at different noise levels may be required for other dose levels. However, they suggest that transfer learning or domain adaptation techniques could be employed to address this limitation.

Conclusion

The paper concludes by emphasizing the introduction of a deep CNN framework designed for low-dose CT reconstruction, combining a deep convolutional neural network with a directional wavelet approach. The proposed method demonstrates greater denoising power for low-dose CT images and offers faster reconstruction times compared to MBIR methods. The success in the 2016 AAPM Low-Dose CT Grand Challenge underscores the effectiveness of the proposed framework and its potential to advance low-dose CT research.