An effective method for classifying brain tumors on MRI images using an enhanced CNN model with LSTM

Early detection of Brain Tumors (BT) is crucial for improving patient survival rates. Due to the difficulties of tumor biopsies, Magnetic Resonance Imaging (MRI) combined with Deep Learning (DL) has become a reliable approach for brain tumor diagnosis. Traditional Machine Learning (MaL) methods depend on hand-crafted features, which are time-consuming, whereas DL automatically extracts powerful features for classification. In this study, a Hybrid Efficient Brain Tumor Classification (HEBTC) model is proposed to classify BTs into three categories: gliomas, meningiomas, and pituitary tumors. The model integrates a modified GoogleNet architecture with Long Short-Term Memory (LSTM) network for improved accuracy. The last five layers of the original GoogleNet are replaced with 14 new layers, incorporating a leaky ReLU activation function to enhance clarity and feature extraction. The model includes two main stages: preprocessing and classification. MRI images are first preprocessed using Gaussian filtering, and then features are extracted through the modified GoogleNet. Classification is performed using LSTM, with its weights optimized by the Adaptive Rider Optimization (ARO) algorithm. Using a publicly available CE-MRI dataset, the proposed HEBTC achieved outstanding results with 99.8% accuracy, 99.6% precision, 99.99% recall, and 99.85% F1-score, outperforming existing Convolutional Neural Networks (CNN) models like ResNet50, MobileNetV2, and AlexNet.