Transfer learning architectures with fine-tuning for brain tumor classification using magnetic resonance imaging

Deep learning methods in artificial intelligence are used for brain tumor diagnosis as they handle a huge amount of data. Compared to computerized tomography (CT), Ultrasound, and X-ray imaging, Magnetic Resonance Imaging (MRI) is effectively used for machine vision-based brain tumor diagnosis. However, due to the complex nature of the brain, brain tumor diagnosis is always challenging. This research aims to study the effectiveness of deep transfer learning architectures in brain tumor diagnosis. This paper applies four transfer learning architectures- InceptionV3, VGG19, DenseNet121, and MobileNet. We used a dataset with data from three benchmark databases of figshare, SARTAJ, and Br35H to validate the models. These databases have four classes: pituitary, no tumor, meningioma, and glioma. Image augmentation is applied to make the classes balanced. Experimental results demonstrate that the MobileNet outperforms competing methods by exhibiting an accuracy of 99.60%. •Classify the brain tumor into four classes using transfer learning and fine-tuning based on magnetic resonance images.•Preprocess and use three benchmark datasets for high accuracy and apply fine-tuning on the transfer learning models.•Modify VGG19, InceptionV3, MobileNet, and DenseNet121 models by adding a single fully connected layer.•Establish standard comparisons between the proposed transfer learning approaches and existing works.•Achieve the best accuracy with MobileNet obtaining 99.60% accuracy in the epoch scenery and InceptionV3 obtaining 98% accuracy in performance.