*2.4. Image Preprocessing*

All axial images were resized to 256 × 256 voxels for neural network training. The axial slices were set to have a distance of 3 mm between each other. The standard deviation and mean values of each image were calculated when retrieved from the database. The image signal intensity was then normalized and applied voxelwise to each image. From all the available mpMRI sequences, only the T2W images were used for training and validation.

### *2.5. Convolutional Neural Network*

The CNN used for this study was a custom modified U-Net. The algorithm's base architecture was derived from a standard U-Net, which is a fully convolutional contracting and expanding architecture [17]. The customized U-Net has a symmetric architecture and uses the same number of layers during subsampling and upsampling. U-Net also employs skip connections that allow the CNN to combine features for the image contraction and expansion pathways. These skip connections enabled the U-Net to use spatial information that could potentially be lost after the image is further downsampled in the contraction pathway. The entire image was trained during a single forward pass and the U-Net classified each image per pixel.

Our customized U-Net was extended to incorporate three dimensions during training and then produce outputs in two dimensions (Figure 1). Five layers were chosen empirically. In each layer, the image was processed by batch normalization, convolution, rectified linear unit (ReLU) activation, and downsampling with strided convolutions by a factor of 2. The 5 layers used 4, 8, 16, 32, and 32 filters per convolution. The image was downsampled until it became a 1 × 1 × 1 matrix before it underwent expansion. During the expansion pathway, the image was upsampled and a skip connection allowed the upsampled image to combine spatial information from the contraction pathway.

**Figure 1.** All neural network runs were completed on a U-Net with 5 layers. The number of channels used were 4, 8, 16, 32, and 32 for the 5 layers.
