Figure 4
From: Computer-aided diagnosis of pectus excavatum using CT images and deep learning methods
Block-wise fine-tuning process of the pretrained VGG network. Block1-Block5 represent the five convolutional blocks of the VGG network. Before fine-tuning, we modified the fully connected layers and built a new SoftMax classifier with three outputs for three categories: normal, mild and severe. B0 fine-tuning means that we freeze all the convolutional blocks and only train the new fully connected layers. Then, we gradually unfreeze the convolutional blocks until reaching Block 1. Bi (where i is equal to 1, 2, 3, or 4) fine-tuning means that we freeze all the pretrained architecture except for the last i block and fully connected layers. Full fine-tuning means that we retrain all the convolutional blocks and fully connected layers.
