Figure 4.

Schematic of the 3D-ResNet-152 that was used for this analysis. Each image was first passed through a 3D convolutional layer (7 × 7 × 7, 64 filters) with ReLu activation and batch normalization, followed by a series of 50 residual units, each with 3 convolutional layers (bottom panel). This output was relayed to a fully-connected dense layer with one output and sigmoid activation (top panel).