Figure 1.

The deep learning model processes a 3D T1-weighted image via a single-input, dual-output 3D convolutional neural network (CNN) to produce estimated multi-label masks for brain tissues (background, white matter, gray matter, cerebrospinal fluid) and brain structures (background + 22 brain structures). The CNN is based on the widely used 3D U-net architecture, which operates on 3D patches of the input scan. Each convolutional layer utilizes $3\times 3\times 3$ kernels, except for the two convolutional layers before the softmax layers, which use $1\times 1\times 1$ kernels. Weight normalization and leaky ReLU (slope = 0.20) are employed. The output patches have dimensions of $88\times 88\times 88$ voxels, which are smaller than the input patches’ dimensions ($128\times 128\times 128$ voxels) due to the use of valid convolutions, mitigating off-patch-center bias.