Comparison of Neural Deblurring Methods for Improving SPECT Reconstruction Quality

Single photon emission computed tomography (SPECT) is a nuclear imaging technique which has numerous applications in the clinical and medical research domain. However, SPECT images suffer from various types of image degradation which may directly influence the quality of the final resulting image. In recent years, due to the fast development of machine learning, many studies have emerged that utilize neural networks (NNs) to improve the image quality of SPECT. Yet, many of of these developing techniques require a significant amount of time and data to train the NN in order to achieve acceptable performance. In this work, we investigated several NN deblurring models which have been pretrained but were never previously trained for medical imaging tasks or with SPECT reconstruction. These methods provide a fast approach for deblurring since there is no extra training work required. Visually inspecting all the methods demonstrated each method's ability to effectively reduce the amount of blur. We also used a quantitative measurement called Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) to evaluate the output image quality of those four methods. The result shows all the neural deblurring methods such as SwinIR and ESRGAN are able to effectively produce a clearer and more readable image from the blurry SPECT reconstructions.