While low-dose CT confers a significantly reduced radiation risk, a consequence of lowering the X-ray flux comes in a corresponding reduction in image quality. A number of different approaches have been proposed to compensate for this, including sinogram filtering, iterative image reconstruction and image post-processing. The first two approaches are difficult to implement, relying as they do on raw projection data, which is usually not made available by commercial CT vendors. While image processing avoids this issue, it is challenging to optimally filter out the non-uniform noise found in low-dose CT images.

A different approach may lie in the use of deep learning, which can learn high-level image features from pixel-level data. A convolutional neural network is currently a machine-learning approach that uses a multi-layered artificial neural network, the design of which is inspired by the inner-working process of the visual cortex. This approach has recently been shown to be successful at a variety of complicated tasks, from picture recognition to speech translation and even automobile driving.

In their study, computer scientist Yi Zhang of Sichuan University, CT expert Ge Wang of Rensselaer Polytechnic Institute, and colleagues have applied a convolutional neural network to process noisy CT images, with the stacked network layers allowing for increased processing power and a better balance between noise removal and information conservation.

The approach performs three tasks. First, individual patches are encoded and embedded into a feature space; they are also non-linearly filtered. Finally, the processed overlapping patches are merged to form a final image. Training samples are used to teach the neural network, and as it works, the network learns – from the data itself, rather than prior assumptions about the noise characteristics – leading to a smart mapping between the noisy low-dose scan and its clean, normal-dose equivalent.

"Such a deep network allows us to divide and conquer; that is, we can handle a complex situation step by step," explained Zhang. "Normally, each layer integrates multiple inputs linearly and then computes a non-linear transformation of the weighted sum."

Performance comparison

To test the network, the researchers used normal-dose CT images from the National Cancer Imaging Archive, from which they artificially generated corresponding low-dose images by applying Poisson noise. They also examined low-dose CT scans of a sheep model.

The convolutional neural network approach was compared with three existing methods – ASD-POCS (an early iterative CT reconstruction method), K-SVD (a dictionary learning method) and BM3D (a popular de-noising method based on the self-similarity of images) – and was found to outperform each of the alternatives in terms of peak signal-to-noise ratio, root mean square error and structural similarity index. It also operated an order of magnitude faster than iterative reconstruction methods.

The convolutional neural network approach would be "technically easy" to apply in a clinical setting, Zhang said, and has no requirement to access raw projection data from CT scanners.

Michiel Kallenberg – a medical physicist from the University of Copenhagen who was also not involved in this study – calls the new noise-reduction method interesting. "The authors are amongst the first to show that deep learning has the potential to play a key role in noise reduction in low-dose CT."

"This work provides promising results in reducing the noise of low-dose CT images, which is an important issue in medical physics," agreed Eunhee Kang, a medical physicist at the Korea Advanced Institute of Science and Technology who was not involved in this study. "The era for deep learning has finally come to image reconstruction and medical physics."

Kang cautioned, however, that the work has several limitations. "First, this network structure is not considered 'deep learning', because it has only three layers. This type of network is often referred to as a 'flat' network, and is considered inferior," she explained. In their paper, however, the authors justify their reduced-layers approach on the grounds of the reduced computational cost afforded by such – although they plan to investigate the performance of deeper networks with larger capacities in the future.

Alongside this, the researchers are also looking to extend their approach to other medical imaging problems, including such challenges as sparse-view reconstruction, interior tomography, limited-view reconstruction and metal artefact reduction.

Related stories

• Neural network cuts false-positive recalls
• CT lung dose cut to 0.25 mSv
• Omni-tomography: the ultimate in multimodality imaging?
• Iterative reconstruction lowers CT dose