The application of Deep Learning to improving low count SPECT imaging

Abstract

Preclinical SPECT systems such as the U-SPECT have been able to achieve sub-half-millimetre spatial resolution with the use of cylindrical pinhole collimators. Utilising this type of collimator comes at the cost of a reduction in the total number of detection events that take place. In order to compensate for this either the activity of the radio pharmaceutical must be increased or the exposure time must be extended. Ideally the dose received by a patient or subject is kept to a minimum. The goal of this study was therefore to investigate the application of deep learning to SPECT imaging, specifically to improve low count images to resemble high count images in the projection domain. The projection domain was chosen over the image domain as projections are easily generated in large quantities, while reconstructed images take large amounts of time and computation to generate. A previous BSc Thesis study constructed a neural network (a Perceptual Loss Network) to this end, and it was concluded that the training set was too small and specific for the neural network to be more generally applicable. In this study therefore the training set of the neural network was expanded on with different phantom types such as Derenzo hot rod phantoms, Jaszczak phantoms and uniform phantoms of various shapes and sizes. Phantoms were simulated and measured using the EXIRAD-3D, and projection pairs (low and high count) were generated. Several network architecture improvements were also explored, such as processing the projection images from different detectors together using width concatenation or applying down sampling. Phantom test sets were used in order to determine whether expanding the training set and making adjustments such as width concatenation or applying down sampling had a positive effect on the neural network’s ability to improve varying SPECT projections. The projections of these test sets were improved by the neural networks and then reconstructed to 3D arrays in the image domain. The reconstructions would then be compared against the low count reconstruction as well as those produced by the original neural network. It became apparent that the neural networks do not perform well on very low count projections. It is assumed that this is because the projections have too little information contained within them for the neural networks to determine how to improve them. It may be possible to improve the neural networks by expanding the training set further with very low count projections. The quality of the reconstructions was determined quantitatively using the Contrast to Noise Ratio (CNR) for Derenzo type phantoms and uniformity for uniform type phantoms. Expanding the training set showed slight improvement in reconstruction CNR but was not considered significantly better. Applying width concatenation as well as expanding the training set seemed to improve results further, but the increase in resource requirements and computing time may not be justified for the marginal increase in CNR. Expanding the training set and applying down sampling proved to be very promising, increasing the CNR from anywhere between 0.35 to upwards of 0.75 in some cases. It also showed the most potential when it came to improving physically measured SPECT projections. The uniform phantoms had varying results. Very low count uniform cylinder phantoms were able to be improved by neural networks, but did not seem to benefit much from training on the expanded training set. It also seemed that trying to improve higher count uniform cylinder phantoms was difficult for the neural networks as there seem to be artifacts in the neural network reconstructions that decrease uniformity. It is recommended to further examine and improve the down sampling technique used in this study. It may also be interesting to combine the individual improvements, expanding the training set, using width concatenation and applying down sampling simultaneously, to see whether this can offer a better neural network for improving low count SPECT projections.