U-Net-Based Medical Image Segmentation and Structural Refinement

In this work, we introduce NGU-Net, a new framework based on U-Net that includes two main modifications: (1) batch normalization applied after every second convolutional layer to reduce the variation of internal covariates and increase the stability of training, and (2) a custom activation layer that replaces the second convolutional layer to better capture meaningful features. These modifications lead to faster convergence during training and significant improvements in segmentation accuracy, as demonstrated in brain MRI experiments.