Abstract <p>Brain tumors, with their various features and potential for rapid growth and development, offer significant challenges in healthcare. The accurate classification of these tumors by MRI (magnetic resonance imaging) is important for several reasons. Different tumor types have diverse features, and correct &amp; early classification is important for treatment planning and improving patient survival rates. Traditionally, this classification relied on expert neurologists analyzing MRI scans, a process prone to subjectivity and human error. Accurately classify multi-class brain tumors using MR images is a challenging issue due to different tumor characteristics. Recently, deep learning techniques achieve precise results in disease classification including in brain tumors. The aim of this work is to design a patch-based deep learning model to accurately classify brain tumors. In this research, EfficientNet based MBconv (Mobile inverted bottleneck convolution) Blocks with deep supervision mechanism (DSM) is used to achieve high accuracy and reduce true-negative and false-positive rate. DSM uses a dilated convolutional block with both adjacent &amp; overlapping patches to extract both local and global features. Dilation effectively enlarges the receptive field of filters, enabling them to capture more contextual information without affecting the number of parameters or computational time. Proposed model is evaluated using multiple dataset split, five-fold, and leave-2-out cross-validation on 3 distinct datasets. Proposed model achieves 0.98 accuracy, 0.97 Recall &amp; 0.97 Precision, 0.99 accuracy, 0.97 Recall &amp; 0.98 Precision, and 0.99 accuracy, 0.96 Recall &amp; 0.98 Precision on Figshare, Kaggle, and Sartaj datasets respectively.</p> Graphical abstract <p></p>