Figure 3.

An overview of the deep learning pipeline for prognosis prediction. Patches of size 224 × 224 × 3 are randomly sampled from whole slide images at a 10 × magnification level. The ResNet-18 Convolutional Neural Network transformed each patch into a 512 × 1 vector. Average pooling is performed at the patient level. The patient level vectors then go through a two-layer fully connected network with a final output size of 1, which can be interpreted as risk scores. Cox proportional hazards loss is calculated using the risk scores with consideration of follow-up time and vital status. The gradient is calculated and backpropagated through the fully connected layers and the ResNet-18 layers to train the entire model.