Tumour substructures’ segmentation is an essential practice for treatment planning and surgical procedure. Manual segmentation is a critical and time-consuming process due to the size of magnetic resonance imaging (MRI). In recent years, automatic methods have been proposed, along with evolving computing techniques and modern architecture, to detect the uncertainty of the brain. We suggested a very efficient modified U-Network (EMU-Net) design to segment the tumour substructures effectively. It was organized into two phases: classification and segmentation using a convolutional neural network (CNN) and modified U-Net architecture on the BraTS 2020 dataset. The classification phase helps find tumorous images from the dataset, enhancing segmentation performance. It contains two independent types (Type-A and Type-B) of classifications using GLCM and CNN, respectively. Based on the classification accuracy of Type-A and Type-B, optimal results are produced to test the segmentation phase. The segmentation modification denotes converting low-level features to high levels by adding skip and cross-connections in each convolutional layer. Real-time datasets were collected from medical centres, which will help test the segmentation’s accuracy. Type-B performs with 99.56% accuracy on tumour image classification during classification. Segmentation result by EMU-Net is achieved against the tumour core (TC), whole tumour (WT), and enhanced tumour (ET) at 85%, 91%, and 79%, respectively. In the end, the 3D tumour volume of substructures is constructed from the obtained results.