Enhanced MRI-Based Brain Tumor Detection Using Deep Learning Architectures and Optimized Machine Learning Techniques

Early brain tumour detection is essential for better treatment results. Biopsies taken during invasive brain procedures are traditionally used to diagnose brain tumours. Tumour identification and classification can be done noninvasively with computational intelligence approaches. This work uses MRI images to diagnose pituitary gland tumours, gliomas, meningiomas, and normal brain function using two deep learning approaches and six data mining techniques. To improve model performance, preprocessing and augmentation were applied to a dataset of 3000 MRI images. Optimised architectures with numerous convolutional in nature, pooling, and batch normalisation layers were used to create an efficient convolutional auto-encoder network and a unique 2D CNN. The 2D CNN outperformed the auto-encoder model with $95.63 \%$ accuracy as well as $94 \%$ recall, achieving a training precision of $96.47 \%$ and average recall of $95 \%$. For both networks, the ROC-AUC values were close to 1. K-Nearest Neighbours was the most accurate machine learning technique ($86 \%$), while Multilayer perception was the least effective ($28 \%$). The suggested deep learning models outperformed machine learning techniques, according to statistical assessments (p-value < 0.05). These results demonstrate how sophisticated deep learning frameworks may be used to accurately detect brain tumours early on.