Deep Learning Features Restoration and Regional Longitudinal Fitting of Computed Tomography Images using Convolution Neural Network

Abstract

Positron Emission Tomography (PET) is suggested for its high potential Deep Learning (DL) diagnostic imaging with a profound learning approach. The network training is done using clear images but reconstructing the low resolution images using Poisson operation. In training the Convolutional Neural Networks (CNN) at a default noise level, a major issue for their generic applicability is the noise level discrepancy. The noise level varies considerably in each iteration reduces the overall efficiency. The results and measured efficiency loss in different noise environments with various noise levels due to inadequate current trials is also presented. To fix this problem, a local linear fitting function is represented before improving the image quality. It indicates that the resulting approach is resilient to noise levels despite the network being educated at a fixed noise level. The proposed protocol is demonstrated to exceed traditional approaches based on total variance and penalty by mean and standard deviation via simulations and trials.