Figure 4.
The 2-stage nnU-Net deep learning architecture. This contained a segmentation network and a category label calculation module. Preprocessing was performed before model training. In the training procedure, the model was trained with a combination of dice and cross-entropy loss. Data augmentation was performed to avoid overfitting. Connected component analysis was used as the postprocessing technique. After inference, each voxel point obtained a predictive probability for the classification. The category label of the whole image was determined according to the maximum number of voxels in one category.