CT-image Super Resolution Using 3D Convolutional Neural Network

Abstract: Computed Tomography (CT) imaging technique is widely used in geological exploration, medical diagnosis and other fields. In practice, however, the resolution of CT image is usually limited by scanning devices and great expense. Super resolution (SR) methods based on deep learning have achieved surprising performance in two-dimensional (2D) images. Unfortunately, there are few effective SR algorithms for three-dimensional (3D) images. In this paper, we proposed a novel network named as three-dimensional super resolution convolutional neural network (3DSRCNN) to realize voxel super resolution for CT images. To solve the practical problems in training process such as slow convergence of network training, insufficient memory, etc., we utilized adjustable learning rate, residual-learning, gradient clipping, momentum stochastic gradient descent (SGD) strategies to optimize training procedure. In addition, we have explored the empirical guidelines to set appropriate number of layers of network and how to use residual learning strategy. Additionally, previous learning-based algorithms need to separately train for different scale factors for reconstruction, yet our single model can complete the multi-scale SR. At last, our method has better performance in terms of PSNR, SSIM and efficiency compared with conventional methods.

• The paper introduces 3DSRCNN, a 12-layer 3D CNN that significantly boosts CT image resolution, achieving an average PSNR increase of 4.72dB and SSIM improvement of 0.037.
• It leverages advanced techniques such as residual learning, gradient clipping, and momentum SGD to address challenges like slow convergence and high memory usage in volumetric data.
• Experimental results highlight an optimal network depth for SR, offering practical benefits for cost-effective, high-quality CT imaging in medical and geoscientific applications.

CT-image Super Resolution Using 3D Convolutional Neural Network

Introduction

The paper "CT-image Super Resolution Using 3D Convolutional Neural Network" (1806.09074) introduces a novel approach for enhancing the resolution of CT images through a 3D convolutional neural network (CNN) architecture. Given the inherent limitations and cost constraints of CT scanning devices, obtaining high-resolution 3D CT images is challenging. The 3DSRCNN, proposed in this research, addresses this challenge by enabling super-resolution (SR) of voxel-based images, transcending the traditional focus on 2D image SR. The proposed network demonstrates significant improvements in PSNR and SSIM when compared to conventional methods, while effectively managing computational intensity and memory consumption inherent to 3D image processing.

Methodology

The 3DSRCNN is a 12-layer 3D convolutional network designed to leverage the spatial information in volumetric CT images. Each convolutional layer comprises 64 filters, which are instrumental in capturing the diverse spatial patterns essential for the SR task. The network is fortified with strategies like residual learning, gradient clipping, adjustable learning rate, and momentum stochastic gradient descent (SGD) to optimize the training process. These strategies counteract slow convergence, gradient instability, and high memory consumption, which are common challenges in training deeper networks on 3D data.

Prior to network training, input data is pre-processed by downsampling ground truth CT images to create multi-scale low-resolution images, which are then cropped into manageable sub-blocks. This preprocessing step not only aids in pattern recognition across scales but also mitigates memory overload by limiting the input block size.

Experimental Results

The research evaluates the 3DSRCNN against several state-of-the-art methods using PSNR and SSIM as metrics for image quality. The model exhibits superior performance, with an average increase of 4.72dB in PSNR and 0.037 in SSIM compared to low-resolution input images. A notable feature of the proposed network is its multi-scale capability—a single model is trained to handle multiple upscaling factors (×2, ×3, ×4), thereby reducing the need for separate models for different resolutions. This attribute is critical for practical applications, reducing both storage and computational costs.

The experiments conducted reveal that the network gains in PSNR until the depth reaches 12 layers, beyond which additional layers do not significantly enhance performance. This indicates that there is an optimal trade-off between network depth and the effectiveness of SR, informed by the dataset characteristics and computational resources.

Implications and Future Work

The findings from this study have substantial implications for enhancing CT imaging, particularly in fields where obtaining high-resolution data is either costly or physically impractical. The ability to accurately reconstruct high-resolution images from low-resolution scans can improve diagnostic and analytical capabilities in medical and geoscientific applications. Additionally, the use of 3DSRCNN in SR tasks can be extended to other volumetric imaging modalities, potentially improving image quality across diverse fields.

Future directions for this research involve theoretical exploration of network depth effects on SR performance and developing more efficient training techniques to manage larger datasets. Such advancements could further reduce training times and enhance the applicability of this technology to real-world scenarios involving extensive 3D data.

Conclusion

This paper successfully presents a compelling advancement in 3D CT image super-resolution through the 3DSRCNN. By optimizing the network architecture and training procedures, the research bridges a critical gap in voxel-based image enhancement. The significance of this research lies in its practical application potential and its contribution to the broader domain of image processing and deep learning. Future research is poised to build upon these findings, paving the way for further expansions in volumetric image processing advancements.