Brain Tumor Detection Using MRI Images
TL;DR
The study adds a rigorous benchmarking mechanism and empirical evidence for adopting ResNet50 as a robust model for multi-class brain tumour diagnosis and highlights the power of deep residual learning for solving some of the difficulties in classifying brain MRI, such as inter-class similarity and feature heterogeneity.
Abstract
Brain tumors represent an important clinical disease of the nervous system that requires early and accurate diagnosis for improved patient prognoses. Magnetic Resonance Imaging (MRI) is still the modality of choice for the evaluation of brain tumor due to its better soft tissue contrast and non-invasive nature. Nevertheless, manual interpretation of MRI examinations is difficult and subject to inter-observer variation. The present study aims at a thorough comparative evaluation of deep learning architectures for fully automated, multi-class brain tumor classification based on MRI data. A unified experimental framework is designed to empirically evaluate a custom convolutional neural network (CNNs) trained with the scratch method and four pre-trained transfer learning methods, including VGG16, ResNet50, DenseNet121, and MobileNetV3. The models are trained and validated using a large, consolidated data set of four classes: glioma, meningioma, pituitary tumour and healthy brain images. In order to make sure the comparability between models is true, same preprocessing, data augmentation and dataset partition strategy and the evaluation measure are followed by all the models. Performance is measured in terms of accuracy, precision, recall, F1 score, confusion matrices and roc-auc analysis. Results show that transfer learning models significantly exceed the custom CNN. Among the architectures evaluated, ResNet50 can achieve the best classification performance, test accuracy is 98.38% and macro-averaged F1-score is 0.9837, which can be regarded as good classification and generalisability. These results highlight the power of deep residual learning for solving some of the difficulties in classifying brain MRI, such as inter-class similarity and feature heterogeneity. The study adds a rigorous benchmarking mechanism and empirical evidence for adopting ResNet50 as a robust model for multi-class brain tumour diagnosis. Future research may extend this work by develop a web-based application that enables users to upload MRI images for automated brain tumor detection and classification.