Skip to main content

Low-dose Computed Tomography for pediatric applications

Periodic Reporting for period 2 - LowD-CT (Low-dose Computed Tomography for pediatric applications)

Okres sprawozdawczy: 2020-03-25 do 2021-03-24

X-ray computed tomography (CT) imaging is an important medical diagnostic technique where a rotating X-ray source and detector are used to image the patient from all directions, allowing a 3D map of the tissue density to be reconstructed. This imaging modality is widely used in the clinic with an estimated 85 million examinations per year in the USA alone in 2012. In children, CT examinations are used for diagnosing congenital heart disease, finding tumors, identifying post-traumatic changes in the brain and planning surgery, among other applications. However, the radiation dose must be kept as low as reasonably achievable to minimize the risk of cancer, in particular for children. One way to lower the radiation dose is to employ photon-counting detectors. This new detector type promises several improvements: lower dose, higher spatial resolution and better measurement accuracy in the images, which is expected to translate to improved diagnostic accuracy and lower cancer risk for the patient. Also, the ability of this detector type to remove beam-hardening artifacts allows them to be used with low X-ray tube voltages (70-100 kV), which is preferable when imaging small patients. These advantages are particularly important for imaging of small children.
The purpose of this project is to use the photon-counting silicon-strip detector with a new reconstruction algorithm, combining fast reconstruction time with optimal dose utilization, to show a proof of principle for a new technique: low-dose photon counting CT. To achieve this, the project has the following specific objectives:
Objective 1 is to develop a computationally inexpensive reconstruction algorithm for low-dose photon counting CT that makes optimal use of the measured data in terms of low-contrast detectability, spatial resolution and accuracy of the reconstructed CT numbers.
Objective 2 is to compare the image quality of low-dose photon-counting CT images to a state-of-the-art CT scanner in clinical use, at equal dose.
Objective 3 is to compare the detectability in photon-counting low-dose CT images of phantoms to the detectability of state-of-the-art x-ray radiography images at equal dose.
However, when carrying out the action, it was decided to change objective 3 to "Develop a deep-learning based image reconstruction algorithm for material-selective photon-counting imaging and evaluate this method in a simulation study."
During the first year of the project the following was performed:
1. A simulation model of a photon-counting CT system was developed. This simulation model is based on an existing software package for CT simulations (CatSim), which has been extended to include a model of a photon-counting detector. This model incorporates physical effects such as the spectral response of a photon-counting detector, crosstalk between different detector pixels, scatter from the imaged object, and geometric blur from the x-ray source.
2. A new hybrid iterative-analytical reconstruction algorithm has been developed. This algorithm builds on first solving a simpler reconstruction problem iteratively and then adding correction terms for incorporating additional physical effects. This new method has been tested on simulated head CT images, and the results show that two correction terms are enough to give a very small error (around 0.2 Hounsfield units or less) relative to a fully converged solution and preliminary results show that the new method can be applied successfully to basis material decomposition of measurements in spectral CT.
After the start of the second (final) year of the project the following tasks were performed:
3. A deep-learning image reconstruction algorithm was developed, through a collaboration with Alma Eguizabal in the Department of Mathematics, KTH. We generated a training set of material-selective CT images and train a deep neural network to denoise the projection CT data. The results are extremely promising with a very substantial improvement in image quality compared to conventional spectral photon-counting CT image reconstruction (material decomposition followed by filtered backprojection).
4. Through the MSc thesis of Emanuel Ström, under the supervision of the experienced researcher together with Ozan Öktem at the department of Mathematics at KTH, it was demonstrated that the complete forward model modelling the physics of scattering in the x-ray detector can be approximated by a simplified forward model followed by a learned correction implemented as a deep neural network. The results show that replacing an approximate model with an approach that combines physics modelling with deep learning can decrease the reconstruction error by an order of magnitude while keeping the computation time at a comparable level.
The results have been presented at the SPIE Medical Imaging 2020 conference, published in Proc. SPIE 113121H (2020), The Lindau Online Science Days 2020 (Jun 28 – July 1, 2020), SPIE Medical Imaging 2021 (Proc. SPIE 115954G and 1159546, 2021) and several research seminars. In addition, a review article on photon-counting CT has been published (Phys. Med. Biol., 66(3) 03TR01, 2021) and the research has been communicated to a wider audience through an exhibit at Swedish Museum of Science and Technology and a popular-science lecture for undergraduates.
Two provisional US patent applications were filed based on this research. These are assigned to Prismatic Sensors AB which was acquired by GE Healthcare in 2020, thus increasing the chances of future commercial exploitation of the results.
The expected long-term impact of this project is a new reconstruction algorithm which makes optimal use of low-dose photon-counting CT image data and is fast enough to be integrated into the clinical workflow. Based on previous studies, photon-counting CT is expected to allow a dose reduction by at least a factor of 2 without compromising diagnostic quality and can enable even larger dose savings with further detector improvements in the future. This will have a positive impact on children’s health by decreasing the radiation exposure and enabling new examinations that are not done today because of radiation safety concerns. Apart from reducing dose, the new hybrid algorithm can also be applied to realize other potential benefits of photon-counting CT, such as higher spatial resolution and quantitative imaging, and therefore promises to open up a new field of research into CT technology improvements.
Even larger potential impact is enabled by deep-learning based image reconstruction methods, demonstrating that combining physics-based modeling with deep learning is a very promising research field in this area. Our results show a very substantial reduction in noise compared to existing image reconstruction methods, suggesting that a future introduction of photon-counting CT with deep-learning reconstruction in routine clinical use can lead to new and improved diagnosis methods, ultimately saving human lives, including but not limited to children.
In addition, the project has already had societal impact its close connection with Prismatic Sensors AB, a startup company commercializing the photon-counting spectral CT technology that was acquired by GE Healthcare in December 2020. This project and the intellectual property generated through it, including two patent applications, has therefore already contributed to hi-tech jobs and industrial competitiveness in Europe.