Comparative Study of Customized CNN Model and Transfer Learning for Brain Tumor Classification

Brain tumors are abnormal cell growths in the brain, which can be cancerous or non-cancerous. They have diverse effects, including cognitive and neurological impairments, as well as personality and behavioral changes. This study proposes a multiclass classification framework utilizing a plain CNN architecture and Transfer learning to accurately identify brain tumor types: meningioma, glioma, pituitary tumor, and no tumor. We have explored transfer learning models such as VGG16, AlexNet, and ResNet50, these are initialized with pre-trained weights and fine-tuned on a comprehensive brain tumor dataset which has an equal distribution among classes. Optimizers such as Adam, AdaDelta, and SGD are modified to enhance performance. AdaDelta was used as the VGG16 optimizer, and it helped us achieve the highest accuracy of 99.83% with a loss of 0.74%. ResNet50 on the other hand, had the lowest accuracy 70% and highest loss 66.68%. While AlexNet without dropouts and AdaDelta as optimizer gives highest loss of 92.83%. The custom CNN model's loss was 0.04%, which is the lowest of any of them. The findings contribute to robust tools for accurate brain tumor diagnosis and treatment planning.