An Efficient Brain Tumor Classification Framework using Multiscale Shuffle Attention Net and Adaptive Trans-SegUnet for Segmentation

Automated brain tumors diagnosis from MRI images has been profoundly researched for the past few years. Evaluating an MRI scan image is a skilled, tedious, and challenging operation. Recently, many segmentation approaches have been defined for identifying the target object from highly complex MRI images, despite failing to address the time complexity and design issues, which are critical for the medical field for timely diagnosis. Hence, this project aims to execute an effective system using images from MRI scans. In this model development, the first step is to obtain the essential MRI images of brain tumors using standard sources. The segmentation phase is applied to the obtained images. Here, the Adaptive Trans-SegUNet (A-Trans-SegUNet) model is developed to effectively segment the gathered MRI images, and the proposed Improved Wild Horse Optimizer (IWHO) is used to modify the parameters of the A-Trans-SegUNet model effectively. Afterwards, the brain tumor classification step is executed. During this stage, the Multiscale Shuffle Attention Network (M-SANet) is developed to categorize the type of brain tumors according to classes related to different tumor types. Simulations are carried out to show a higher percentage of accuracy while identification and classification by the recommended segmentation and classification technique, when compared with other methods. The deployed M-SANet model achieved a classification accuracy of over 7.23%, 5.13%, 3.2%, and 1.15%, more accurate than the DenseNet, MobileNet, Inception, and ResNet models, showing the effectiveness in brain tumor classification.