Enhancing Brain Tumor Detection in MRI Images through Transfer Learning and Multi Stage Deep CNN Models: A Comparative Study

Brain tumors are abnormal growth of cells which are very serious issue, and the early diagnosis is the main factor for the enhancement of the patients' situation. The traditional methods of brain tumor detection and classification are usually time-consuming, and often the human error is involved. Also, they cannot at all accurately classify the different types of tumors. The deep learning models mainly CNNs have been the main factors for the significant transformation of the medical image analysis. CNNs are applied to automatically extract the complex features from MRI images which make it possible to get the advantages of the exact identification of the tumor boundaries and the slight variations that can be ignored by human. This work investigates the use of transfer learning technique to improve the accuracy of CNN models for brain tumor detection and classification. We tested VGG16 and ResNet50 models on 4000 MRI images with tumor and no-tumor. Both scratch and pre-trained models were used. Results show that pre-trained VGG16 and ResNet50 achieved 99.44% and 99.55% accuracy while scratch models achieved 59.64% and 96.02% accuracy respectively. This result proves the power of transfer learning technique in deep learning models especially CNNs for brain tumor detection and classification.