[Paper Review] Can Deep Learning Outperform Modern Commercial CT Image Reconstruction Methods?

Hongming Shan, Atul Padole+6 more|arXiv (Cornell University)|Nov 8, 2018

Medical Imaging Techniques and Applications1 references71 citations

TL;DR

The paper presents a progressive deep learning denoising network (MAP-NN) for low-dose CT that competes with commercial iterative reconstructions across vendors and regions, offering faster performance and radiologist-in-the-loop optimization.

ABSTRACT

Commercial iterative reconstruction techniques on modern CT scanners target radiation dose reduction but there are lingering concerns over their impact on image appearance and low contrast detectability. Recently, machine learning, especially deep learning, has been actively investigated for CT. Here we design a novel neural network architecture for low-dose CT (LDCT) and compare it with commercial iterative reconstruction methods used for standard of care CT. While popular neural networks are trained for end-to-end mapping, driven by big data, our novel neural network is intended for end-to-process mapping so that intermediate image targets are obtained with the associated search gradients along which the final image targets are gradually reached. This learned dynamic process allows to include radiologists in the training loop to optimize the LDCT denoising workflow in a task-specific fashion with the denoising depth as a key parameter. Our progressive denoising network was trained with the Mayo LDCT Challenge Dataset, and tested on images of the chest and abdominal regions scanned on the CT scanners made by three leading CT vendors. The best deep learning based reconstructions are systematically compared to the best iterative reconstructions in a double-blinded reader study. It is found that our deep learning approach performs either comparably or favorably in terms of noise suppression and structural fidelity, and runs orders of magnitude faster than the commercial iterative CT reconstruction algorithms.

Motivation & Objective

• Motivate the use of deep learning to improve low-dose CT image quality while reducing radiation dose.
• Introduce a progressive denoising network that integrates radiologists in the training loop.
• Compare the MAP-NN approach to three vendors’ commercial iterative reconstruction methods across chest and abdomen.

Proposed method

• Propose MAP-NN, a modular progressive denoising network built from CLONE modules for LDCT denoising.
• Train MAP-NN with a loss function combining adversarial loss, mean-squared error, and edge incoherence.
• Share parameters across CLONE modules and evaluate the mapping depth (number of denoising steps).
• Evaluate DL reconstructions against vendor IR methods using a blinded radiologist reader study.
• Use LDCT inputs and compare to NDCT references to assess structural fidelity and noise suppression.

Experimental results

Research questions

• RQ1 Can the MAP-NN progressive denoising approach achieve comparable or better structural fidelity than commercial IR methods for LDCT?
• RQ2 Does MAP-NN offer better noise suppression while maintaining structural fidelity across multiple CT vendors and body regions?
• RQ3 Is the DL approach computationally more efficient than IR methods after training?
• RQ4 Does incorporating radiologists in the loop improve task-specific image quality optimization?

Key findings

• MAP-NN achieves comparable or better structural fidelity and noise suppression versus commercial IR methods across vendors A, B, and C.
• DL reconstructions are rated better or comparable to IR reconstructions for abdomen and chest imaging by radiologists in most cases.
• MAP-NN demonstrates significantly faster run-time performance than IR methods after training (approximately 100 slices per second per mapping depth).
• Best DL reconstructions are preferred over best IR reconstructions for abdominal imaging in vendors A and B, and are comparable for vendor C; for chest, DL is comparable or better in most cases.
• DL approach provides vendor-agnostic image appearance, potentially aiding large-scale radiomics studies.