Gtx 1080

We used an Oculus Rift CV1 with two Oculus sensors and an Asus laptop with an Intel i7 processor, 16 Gb of RAM memory, and a Nvidia GTX 1080 graphics card. The researcher started the experience by pressing a key on the keyboard and had the option to stop it at any time.

A three-class UNet model¹⁴ was trained based on annotation of nuclei centers, nuclei contours, and background. The neural network is comprised of 4 layers and 80 input features. Training was performed using a batch size of 32 with the Adam Optimizer and a learning rate of 0.00005 with a decay rate of 0.98 every 5000 steps until there was no improvement in accuracy or ~100 epochs had been reached. Batch normalization was used to improve training speed. During training, the bottom layer had a dropout rate of 0.35, and L1 regularization was implemented to minimize overfitting^{44 ,45} and early stopping. Training was performed on workstations equipped with NVidia GTX 1080 or NVidia TitanX GPUs.

The 100 preprocessed HCP subjects were split into a training and testing set with a commonly used 80–20% ratio, respectively. The concatenated input had a total of 8 channels which include a channel for T1, a channel for b0, and 6 channels for minimally evenly distributed diffusion gradient directions. The patch size of the input was set to $64\times 64\times 64$ with a $32\times 32\times 32$ overlap. We discarded any patches containing only the background region to reduce the computational cost during the training phase. The HADTI-Net was trained for 100 epochs using Adam with an initial learning rate of 0.0001 and batch size of 8 on a single NVIDIA GTX 1080. For inference, the averages from different overlapping regions between adjacent patches were calculated to reconstruct the final enhanced LAR-DTI volume. The code was implemented on PyTorch (version 1.10.1) and Numpy (version 1.21.2).

All neural network models were trained in the open-source package PyTorch⁷¹ on four NVIDIA GTX 1080 or Titan X graphics processing units. Our optimized model used a simple convolutional architecture for image classification, consisting of alternating (3 × 3) kernels of stride 1 and padding 1 followed by max pooling (Fig. 3a), followed by two fully connected hidden layers (512 and 100 neurons) and rectified linear units as the nonlinear activation function. All neural network models were trained using backpropagation. The optimized training procedure used the Adam⁷² optimizer with a multi-label soft margin loss function with weight decay (L2 penalty, 0.008) and dropout (probability 0.5 for the first two fully connected layers and probability 0.2 for all convolutional layers). Training proceeded with mini-batches of 64 images with real-time data augmentation including random flips, rotations, zoom, shear, and color jitter. When calculating the classification accuracy, a threshold of 0.91, 0.1, and 0.85 was used for cored plaque, diffuse plaque, and CAA prediction, respectively. Predictions with confidence above the threshold were considered to be positives.