PET Image Reconstruction Using a GRU-Convolutional Network

Abstract

Positron emission tomography is widely used for tumor detection and treatment monitoring in oncology. However, the quality of the images depends, among other factors, on the amount of radiopharmaceutical ingested by the patient. In this sense, the quality suffers degradation because there is a limit on the amount of radiation the patient can tolerate. Because of this, image reconstruction algorithms are required to generate images of adequate quality even if the amount of radiopharmaceutical to produce the image is small. In this study, a reconstruction algorithm is proposed based on deep learning using a GRU recurrent network which is expected to model the series of projections produced by the PET scanner as an input sequence to the recurrent network and is capable of reconstructing an image even with low amounts of the radiopharmaceutical. In comparisons using image quality metrics, our proposal achieves a SIMM of 0.95, outperforming other state-of-theart methods. Additionally, tests were performed for the evaluation of the task of lesion detection; the proposed method obtained a better contrast of the lesion with a value of 0.54 when using the weber contrast metric, very similar to the ground truth contrast of 0.55.