Devbox

We trained the models by a stochastic gradient descent (SGD) with a momentum of 0.9 and with a minibatch size of 8 to achieve full GPU utilization. As performed in [31 , 35 (link)], we utilized a fixed, tiny learning rate and weight decay because training is highly sensitive to hyperparameters when unnormalized softmax loss is used. We empirically found that a learning rate of 10⁻¹⁰ and a weight decay of 10⁻¹² were optimal for our application to obtain stable training convergence at the cost of convergence speed. Since training losses eventually converged if the models were trained for sufficient period of epochs, all models in this paper were trained for 500 epochs and the last model was selected without a validation phase to evaluate performance on our held-out test subset. All experiments were run on a Devbox (NVIDIA Corp, Santa Clara, CA) containing four TITAN X GPUs with 12GB of memory per GPU [36 ] and using Nvidia-Caffe (version 0.15.14) and Nvidia DIGITS (version 5.1).

We used radon and iradon functions in Matlab 2018a for generating sinograms and obtaining FBP reconstructed images, respectively. We used Keras (version 2.1.1) with a Tensorflow backend (version 1.3.0) as the framework for developing deep learning models, and performed experiments using an NVIDIA Devbox (Santa Clara, CA) equipped with four TITAN X GPUs with 12 GB of memory per GPU.