Advanced deep learning-based brain tumor classification using a novel customized CNN and optimized residual network

The uncontrollable and rapid growth of brain cells can lead to brain tumors. If left untreated, this condition may result in severe health consequences, including death. Accurate detection and classification are the essential steps toward understanding their mechanisms and ensuring effective treatment. Both tasks are challenging, with brain tumor detection being more complex due to variations in tumor size, structure, and location. Many scholars have employed machine learning and deep learning methods for brain tumor detection. Deep learning (DL) methods provide robust solutions for the detection and classification of brain tumors. Large volumes of healthcare imaging data can be analyzed using these techniques to identify and characterize tumors with high accuracy, often surpassing human performance. In this study, we propose two deep learning models, a novel customized Convolutional Neural Network (CNN) and an optimized ResNet101, to classify brain tumor images into four categories: gliomas, pituitary tumors, meningiomas, and no tumor. We used an MRI image dataset from Kaggle, consisting of 3,264 images. We performed five-fold cross-validation on the training and validation set, and a separate test set was used for final evaluation. The average training accuracy across the five-fold was 99.03±0.01% for the novel customized CNN and 99.87±0.03% for optimized ResNet101, and the average validation accuracy was 96.31±0.01% and 97.23±0.03%, respectively. After the cross-validation, the best-performing fold was then selected and evaluated on the test set, achieving training accuracies of 99.05%, 99.91% and testing accuracies of 97.72%, 98.73%, respectively. The optimized ResNet model achieved the highest performance among the two proposed models. Overall, these findings demonstrate the potential of deep learning models in supporting clinical decision-making for brain tumor classification, which may improve survival rates and human health outcomes.