NGU-Net: a modified U-Net with feature refinement block for accurate medical image segmentation

Abstract Accurate segmentation of biomedical images is essential for medical diagnosis and treatment planning, enabling automated and reliable analysis of complex imaging data. Despite the widespread use of U-Net and its variants, challenges remain, including internal covariate shifts and suboptimal feature extraction, which can limit performance on complex datasets. NGU-Net builds on the U-Net encoder–decoder and introduces a Feature Refinement Block (FRB) that keeps the two-convolution structure of the standard block. In each FRB, a 3 × 3 convolution with ReLU is followed by a second 3 × 3 convolution, batch normalization, and a final ReLU. This ordering stabilizes the internal feature distributions and training, with only a negligible increase in parameters. These modifications contribute to faster convergence during training and improved segmentation performance. Experimental results on brain Magnetic Resonance Imaging (MRI) datasets demonstrate that NGU-Net outperforms the original U-Net and several of its variants, achieving higher accuracy while maintaining computational efficiency. The proposed framework provides an effective architecture for brain tumor segmentation, demonstrating promising performance on The Cancer Genome Atlas Lower-Grade Glioma (TCGA-LGG) dataset that warrants further multi-center external validation.