Deep learning-based Convolutional Neural Network (CNN) architectures are commonly used in medical imaging. Medical imaging data is highly imbalanced. A deep learning architecture on its own is prone to overfit. As a result, we need a generalized model to mitigate total risk. This paper assesses a layer normalization (LN) technique in a CNN-based 3D U-Net for faster training and better generalization in medical imaging. Layer Normalization (LN) is mostly used in Natural Language Processing (NLP) tasks such as question-answering, handwriting sequence generation, etc. along with Recurrent Neural Network (RNN). The usage of LN is yet to be studied in case of medical imaging. In this context, we use brain MRI segmentation and train our model with LN and without normalization. We compare both models and our LN-based model gives $$32\%$$32% less validation loss over without normalization-based model. We achieve validation dice scores of unseen input data passes to LN based model of 0.90 ($$7.5\%$$7.5% higher than without normalization) for edema, 0.74 ($$12.5\%$$12.5% higher than without normalization) for non-enhancing tumor and 0.95 ($$1.5\% $$1.5%higher than without normalization) for enhancing tumor.