[Paper Review] A Cascaded Convolutional Neural Network for X-ray Low-dose CT Image Denoising

Dufan Wu, Kyung Sang Kim+2 more|arXiv (Cornell University)|May 11, 2017

Medical Imaging Techniques and Applications22 references57 citations

TL;DR

The paper proposes a cascaded CNN framework where a sequence of CNN denoisers is trained on progressively denoised data to iteratively reduce artifacts in low-dose CT images, improving PSNR/SSIM over single-pass denoisers.

ABSTRACT

Image denoising techniques are essential to reducing noise levels and enhancing diagnosis reliability in low-dose computed tomography (CT). Machine learning based denoising methods have shown great potential in removing the complex and spatial-variant noises in CT images. However, some residue artifacts would appear in the denoised image due to complexity of noises. A cascaded training network was proposed in this work, where the trained CNN was applied on the training dataset to initiate new trainings and remove artifacts induced by denoising. A cascades of convolutional neural networks (CNN) were built iteratively to achieve better performance with simple CNN structures. Experiments were carried out on 2016 Low-dose CT Grand Challenge datasets to evaluate the method's performance.

Motivation & Objective

• Motivate denoising for low-dose CT to preserve diagnostic quality while reducing radiation exposure.
• Propose a cascaded training scheme that refines denoising by learning residuals at each stage.
• Evaluate the cascaded approach against conventional denoisers and a baseline neural network on standard low-dose CT data.

Proposed method

• Use a residue-learning CNN mapping low-dose CT to the difference between low- and normal-dose images.
• Train f1 on xL to xH to obtain xD(1); train f2 to map xD(1) to xD(1)−xH using inputs stacked with xL to preserve information.
• Repeat cascading with the same structure while adjusting input channels; all CNNs share 3x3x64 convolutional kernels with BN and ReLU except the first and last layers.
• Loss function is L2; training uses ADAM with a learning rate of 1e-4 and weight decay 1e-4; training on patches (40x40) from 3,933 slices across 7 patients.
• The cascaded approach relates to residual networks by stacking inputs rather than adding via skip connections; it is conceptually similar to SSDA in gradually refining denoising.

Experimental results

Research questions

• RQ1 Can cascaded CNNs trained on progressively denoised data reduce artifacts better than a single CNN in low-dose CT denoising?
• RQ2 How does the number of cascades affect objective image quality metrics (PSNR, SSIM) compared to conventional methods?
• RQ3 Do cascaded CNNs preserve diagnostically important structures, such as lesions, better than BM3D/WNNM or a plain MLP?

Key findings

• Cascading significantly boosts PSNR for both original denoised and blended results across CNN5, CNN10, and CNN15.
• Blended results (70% denoised, 30% low-dose) show SSIM improvement with more cascades, while original denoised SSIM may saturate or decline.
• CNN cascades outperform BM3D and WNNM in averaged SSIM on 9 slices; CNN15-3 had the best reported SSIM in this comparison.
• MLP cascades improve less than CNN cascades, indicating CNNs better capture CT noise characteristics.
• Qualitative results show cascaded CNNs reduce strip/block artifacts and better preserve lesion visibility compared to single-pass denoising and conventional methods.