This thesis focuses on the development and evaluation of four convolutional neural network models for the accurate and fast diagnosis of brain cancer. Brain cancer is a life-threatening illness that requires precise diagnosis to improve patient outcomes. To address this need, the AlexNet, EfficientNetV2, NASNet, and Xception models were trained to correctly classify glioma, meningioma, pituitary cancers, or the absence of tumors from MRI images of the brain.

The results demonstrate the potential of convolutional neural networks when properly utilized. Even the weakest model achieved an accuracy of 96%, while the best model achieved a remarkable accuracy of 99.5% when tested on previously unseen clinical images. This research offers a promising outlook on the potential of convolutional neural networks, even with limited computational resources and training data. The findings suggest that institutions with adequate resources can develop a product-ready model with close to 100% accuracy, which could significantly enhance the performance and efficiency of medical practitioners by providing an effective second opinion.

However, it is essential to note that interpretability remains a significant challenge. These models cannot make the final diagnosis independently until their interpretability is fully explained. This work highlights the need for further research on the interpretability of convolutional neural networks, as this will be necessary for their implementation in clinical settings.

Overall, this thesis provides an important contribution to the field of medical diagnosis and machine learning, emphasizing the potential of convolutional neural networks to improve the accuracy and efficiency of brain cancer diagnosis.