<p>Glioblastoma (GBM) is the most dangerous and fatal brain tumor, responsible for the demise of more than fifty percent of patients within one to two years of diagnosis. Tumor segmentation from magnetic resonance imaging (MRI) based neuroimages is pivotal for the in-time clinical treatment process and surgical planning of cancer patients. Early diagnosis and analysis of brain cancer can support a decrease in the human mortality rate. However, accurate segmentation is a challenging task owing to the uneven, irregular, and unstructured tumor boundary connections among substructures such as whole tumor (WT, comprising all classes of tumor structures), tumor core (TC, comprising a necrotic core and non-enhancing core), and enhancing tumor (ET). Moreover, manually analyzing these substructures through massive volumetric neuroimages is time-consuming and requires a lot of costs. To do this, we have developed a better and more efficient multi-step 3D U-Net-based cascade method to find important local and global features and train our network to segment tumor substructures well. The non-invasive multimodal brain MR image modalities (T1, T1c, T2, and FLAIR) were utilized. Our proposed architecture consists of DenseNet encoder blocks and ConvNet decoder blocks. Pre-processing techniques such as bias field correction and normalization are also applied to remove the bias field effect, which greatly affects tumor segmentation. The developed network was accessed using Brain Tumor Image Segmentation (BraTS 2018) data, DATASET1, and private datasets DATASET2. The designed CAD system achieves a dice score of 0.921, 0.879, and 0.821 for DATASET1 and 0.899, 0.838, and 0.780 for DATASET2 that correspond to WT, TC, and ET regions respectively. This study reveals that our model outperforms other advanced brain tumor segmentation methods.</p>