Brain Tumour Classification: A Comparative Analysis of CNN Models and Generalisation Across Datasets

In the realm of medical diagnostics, the identification and categorization of brain tumors are critical tasks, and recent advancements in deep learning hold promise for enhancing these processes. Despite improvements in medical imaging technology, accurately diagnosing and classifying the brain tumors continues to pose a considerable challenge due to the complexity and heterogeneity of tumor types. This study addresses this problem by conducting a comparative evaluation of four convolutional neural network (CNN) models—MobileNet, ResNetV2, Inception-v3, and VGG16—for brain tumor image classification. Our approach involves training these models using an extensive brain tumor image dataset that includes four categories: pituitary tumor, meningioma tumor, glioma tumor, and no tumor. We then evaluate their performance in comparison to established pre-trained models like DenseNet, EfficientNet, and AlexNet. To gauge and compare the effectiveness of each model, we utilize metrics such as F1-score, accuracy, recall, precision, and ROC-AUC. Visualizations of accuracy and loss metrics are presented for clarity. The results indicate that while all models show strong performance, certain models exhibit particular strengths in specific classification tasks. For instance, ResNetV2 demonstrated the highest accuracy and generalization capability on external datasets, while MobileNet offered a balance between accuracy and computational efficiency.