Deep Learning in Neuroimaging: A Comparative Analysis of Models for Brain Tumor Classification using MRI Images

Abstract

The precise classification of brain tumors is essential for timely diagnosis and effective treatment planning. This study proposes a deep learning framework for the automatic categorization of three primary brain tumour types: glioma, meningioma, and pituitary tumors using Magnetic Resonance Imaging data. A dataset publicly accessible on Mendeley, containing 6,056 labeled MRI images categorized into three tumor types, was utilized to train and assess three different models: a hybrid CNN-LSTM network, ResNet18, and VGG16. These models were chosen to investigate both temporal-sequential learning (through the CNN-LSTM) and effective convolutional feature extraction using established transfer learning frameworks (ResNet18 and VGG16). The MRI images underwent various pre-processing steps, which included resizing, normalization, and augmentation, to enhance the robustness of the models. Among the models tested, ResNet18 achieved the highest classification Accuracy of 93.50%, Precision of 93.40%, Recall of 93.40% and F1-Score of 93.40% while the CNN-LSTM following at 92.24% Accuracy, 92.00% Precision, 91.80% Recall and 91.80% F1-Score and VGG16 achieves Accuracy of 91.82%, Precision of 91.80%, Recall of 91.40% and F1-Score of 91.40%. ResNet18 demonstrated improved generalization when applied to various tumor classifications. These findings underline the potential of deep learning, especially ResNet18, as a valuable tool to support radiologists in the early and non-invasive identification of brain tumors, showcasing its importance in the progression of automated neuro-oncological diagnosis.