IDENTIFICATION OF BRAIN TUMOR IN MRI LEVERAGING MULTI-SCALE ATTENTION ARCHITECTURE BASED ON EFFICIENTNETB4

Abstract

It takes a long time to diagnose brain tumors (BT), and radiologists must possess a high degree of skill and expertise. The amount of data that has to be handled has increased along with the number of patients, rendering earlier methods both expensive and inefficient. A state-of-the-art approach for brain cancer detection and MRI segmentation is presented in this work by utilizing the EfficientNetB4 DL architecture. The principal data source is the Figshare Brain Tumor Dataset, which contains 3064 contrast-enhanced T1-weighted MRI slices from tumors within the pituitary, meningioma, and glioma areas. To increase the quality of the images and the performance of the models, extensive preprocessing methods were used, such as image scaling, min-max normalization, CLAHE, and Gaussian blur. Better results have been obtained using the EfficientNetB4 model with 80:20 data splits, with accuracy of 99.7%, F1-score 93.6%, recall 91.0%, precision 96.5%, and dice score 93.3%. A comparison with cutting-edge models such as UNet, VNet, and SVM showed how much better the suggested model was in segmentation accuracy and resilience. Through automated MRI image processing, findings demonstrate EfficientNetB4's potential to facilitate precise and effective brain tumor detection.