Residual Neural Networks for the Prediction of the Regularization Parameters in PET Reconstruction

Positron Emission Tomography (PET) is a medical imaging modality relying on numerical methods that integrate the statistical properties of the measurements and prior assumptions about the images. In order to maximize the computed image quality, PET reconstruction algorithms require the setting of hyperparameters that balance data fidelity with regularization. However, their optimal tuning depends on the statistical properties of the raw data and on the clinical objectives. To address this issue, we propose a supervised deep learning strategy based on a residual neural network that takes the raw measured data (sinogram) as input and automatically predicts the optimal value of the regularization parameter of the modified block Sequential Regularized Expectation Maximization (BSREM) algorithm. The proposed strategy is trained on a synthetic dataset consisting of 2D sinograms and their corresponding optimal regularization parameters. Our results demonstrate the feasibility of the approach leading to improved image reconstruction compared to classical manual tuning methods.