Deep-learning data processing for photon-counting CT
||Deep-learning data processing for photon-counting CT|
||Mats Persson <email@example.com>|
||Kungliga Tekniska högskolan|
||2023-04-17 – 2023-11-01|
X-ray computed tomography (CT) imaging is a widely used medical imaging modality with around 100 million examinations per year in the US alone, providing healthcare with an important tool for diagnosis and treatment planning for a wide range of diseases such as stroke, cancer, and cardiovascular disease.
The latest improvement of this technology is photon-counting CT, which is based on a new, energy-resolving x-ray detector type that promises lower dose, higher spatial resolution, and improved tissue characterization. A key problem is to develop data processing algorithms for photon-counting CT that are able to make optimal use of the measured data while being sufficiently fast to be adopted in clinical use. In recent years, deep-learning based image processing methods have been shown to be highly effective for improving image quality for conventional CT imaging while being computationally inexpensive to apply once the training process is complete. In preliminary work we have seen drastic improvements in image quality for denoising and artifact reduction in photon-counting CT.
Our goal is to develop deep-learning-based data processing methods for reduction of artifacts and noise that can be integrated into the image reconstruction process and lead to improved photon-counting CT imaging. We will train convolutional neural networks on simulated CT scans of numerical phantoms generated from anonymized CT image datasets available online, using images corrupted by artifacts and noise as input and corresponding high-fidelity images as labels. Since the images are anonymized, the data will not contain any protected health information. We will evaluate performance by reconstructing images of test objects resembling human patients imaged with a prototype photon-counting CT scanner, based on technology developed in our research group, that is in operation at MedTechLabs at Karolinska University Hospital.
In our previous work, we found that Berzelius is well suited for this project since the default allocation allows us to train efficiently on datasets up to 40 GB and try out multi-GPU implementations for larger datasets. During the previous allocations we obtained very promising results for both image denoising and ring artifact reduction. In this continuation project, we will continue this line of research by 1) investigating more advanced neural networks for denoising, including 3D models and 2) investigating if the techniques we have developed for correcting ring artifacts can also be used to suppress other types of artifacts such as shades and streaks in the image. If we find that more training capacity is needed, we may use our experience with the default allocation to apply for a larger allocation in the future.
The expected outcome is a proof of concept demonstrating that incorporating deep-learning image processing steps in the CT image reconstruction can play an important role in improving the diagnostic quality of photon-counting CT. This will be an important step towards our long-term goal of introducing a novel highly accurate photon-counting CT imaging technique in clinical practice, with large potential benefits to human health.