Deep Learning-Based Compressed Sensing Image Reconstruction Technology

Summary

Combines deep learning with compressed sensing to synthesize high-quality CT images
Does not require pairs of degraded and ground-truth images, unlike conventional CNN-based methods
Capable of reconstructing low-noise, high-quality images from limited CT projection data
We invite you to consider integrating this technology into your CT systems

Technology Overview & Background

To reduce patient radiation exposure, sparse view CT, which reduces the number of CT scans, and low-dose CT, which scans using reduced X-ray doses, are being utilized. However, reconstruction images from such specialized CT scans have been plagued by streak artifacts, resulting in poor image quality. To address this, machine learning-based image quality improvement technologies are being developed. This idea involves training a convolutional neural network using pairs of high-quality reference images obtained from conventional CT scans and low-quality images obtained from special CT scans to create a machine learning model that can “convert” low-quality images into high-quality reference images. However, in general, a large number of degraded images and reference images are required, which is expected to result in high development costs.

To address this issue, Professor Kudo of Tsukuba University has developed a technology that enables the training of a deep learning-based image quality converter at a lower cost. First, this technology possesses a simulation technique that converts data acquired using conventional CT into data acquired using special CT (such as sparse view CT). This eliminate the need to prepare low-quality images for training. Furthermore, this technology creates training data by applying image reconstruction using compressed sensing to the low-quality images generated through conversion. Using machine learning with data pairs of the reconstructed images and their corresponding ground truth images, this technology successfully created a conversion model with excellent image quality improvement effects. This enables the removal of staircase artifacts, preservation of smooth density transitions, and highly high-quality image reconstruction. This technology enables the generation of very high-quality images even with sparse view CT and low-dose CT, and is expected to lead to the development of high-performance CT devices that reduce the burden on patients and examinees. Professor Kudo has independently developed an image reconstruction algorithm using compressed sensing (*), which has been shown to have high reproducibility compared to similar conventional technologies.

Data

As shown in the figure, sparse view CT (64-direction projection data) was acquired, reconstructed using filter-corrected inverse projection (FBP) method and the present method (non-local total variation (TV) + deep learning (convolutional neural network: CNN)), the MSE value and SSIM value were 1022.02 and 0.58, respectively, for the FBP method, whereas the present method yielded 37.68 and 0.90, confirming a significant improvement in image quality. The “original image” in the figure above is an image reconstructed using the FBP method for normal CT imaging.

Patent

WO2021/182103

Researchers & Academic Institution

Hiroyuki Kudo, PhD (Professor,　Faculty of Engineering, Tsukuba University, Japan), et al.

Expectations

We are seeking companies interested in developing and commercializing CT devices (for medical, dental, or industrial non-destructive testing applications) that adopt this technology. Why not introduce high-quality, high-speed, low-dose sparse view CT image reconstruction based on the technology provided by the University of Tsukuba? If you would like more technical details, please watch the lecture video above and contact us. We will arrange a meeting with the developers. Regarding technology adoption, we can discuss contracts (for a fee) for the provision of reference programs or lectures by developers.
In addition to high-performance reconstruction using machine learning, we also offer image reconstruction technologies for special CT applications such as high-performance reconstruction using compressed sensing, sparse view CT, and interior CT. Please also refer to the following pages.

Project ID:DA-02355c