Fig. 1
From: DiseaseNet: a transfer learning approach to noncommunicable disease classification
DiseaseNet architecture and transfer learning scheme. Left: Fully trained CancerNet. The transfer learning process starts with the fully trained CancerNet. Center: Transformation to DiseaseNet from CancerNet. The weights are frozen for the encoder and decoder. The classification layers (top) are replaced with layers with randomly initialized weights. The output (softmax) layer has four nodes representing the four diseases being classified. In the first round of training, only the classifier is trained. Right: Finetuning of DiseaseNet. Weights of the entire model are unfrozen and allowed to train in the last round until convergence