Figure 2.

Workflow of the deep learning model for the prediction of muscle invasiveness status in bladder cancer patients. (A) Collection of the CT images of MIBC and NMIBC. (B) Semiautomatic segmentation of the tumor region. (C) The masked tumor region and the original tumor region were stacked vertically to form the input volume, and the cropped 2-channel input was constructed. (D) The structure of our deep-learning model. The model was constructed on the basis of Filter-guided Pyramid Network (FGP-Net), a novel 3D convolutional network structure that is designed to capture the global feature and the local features simultaneously. (E) Internal and external validation of our model. CT, computed tomography; FC, fully connected layer.