Fig. 2From: Combining deep learning with a kinetic model to predict dynamic PET images and generate parametric imagesThe relation between the feature extraction network and the kinetic model network (pointwise neural network). We obtained each frame’s feature maps using U-Net. All 220 feature maps were then input into a pointwise neural network, which was implemented by a stack of convolutions with a kernel size of 1. Therefore, each voxel was a 220-dim vector that was fed into the kinetic model networkBack to article page