<p>Both adults and children can develop brain tumors, which are among the most frequent and deadliest forms of cancer. The high expense of brain tumor examination equipment and delayed diagnosis are two of the main factors contributing to the rising death rate. The majority of current methods involve machine learning techniques, but these algorithms have drawbacks with low accuracy, high loss, and high computing cost. This research presents a new deep learning-based method for efficient classification and segmentation. Pre-processing methods are used to recover image quality and eliminate extraneous information once input images are first collected from publically accessible datasets. In this case, the suggestion improves the quality of images with the min-max normalization and Gaussian bilateral filter. An optimized YOLO v9 model is also suggested to process brain tumour segmentation, whereas the weight optimization is processed by the chaotic Harris hawks optimization (CHHO) model. The VGG-16 model is applied to cut off the images of the relevant features. In the selection of features, the Chebyshev Coati optimization algorithm (CCOA) is applied in finding the best feature set. On these characteristics, brain tumours are classified into benign and malignant ones. In this paper, a multi-head attention dense network (Multi-head attention-based residual dense network, or MResDensnet) is proposed to ensure the classification of brain tumours. The performance of the suggested technique is examined using two public datasets, BraTS 2019 and 2020. Moreover, the given strategy provides results with an accuracy of 98.51 per cent and 98.65 per cent on brain tumour classification using BraTS 2019 and 2020, respectively.</p>