Transfer learning with pre-trained CNNs for MRI brain tumor multi-classification: A comparative study of VGG16, VGG19, and inception models

• Main Authors: Abd. Aziz, Siti Nurfarahin, Dziyauddin, Rudzidatul Akmam, Mohd. Noor, Norliza
• Format: Conference or Workshop Item
• Published: 2023
• Subjects: T Technology (General)
• Online Access: http://eprints.utm.my/107764/; http://dx.doi.org/10.1109/NBEC58134.2023.10352589
• Institution: Universiti Teknologi Malaysia

In the field of medical imaging and diagnosis, the advancement of machine learning techniques has brought about significant progress. The classification of brain images, particularly in the context of brain tumors, has evolved from traditional methods to more sophisticated approaches like deep learning, specifically Convolutional Neural Networks (CNNs). The features that CNN extracts were significantly impacted by the size of the training dataset. When the training dataset is small, the CNN often overfits. Therefore, Deep CNNs (DCNN) with transfer learning have been created. This study used data augmentation and transfer learning techniques to examine the potential for categorization of brain MR images by pre-trained DCNN VGG-19, VGG-16, and Inception V3 models. Accuracy, recall, precision, and F1 score validation on the test set revealed that the pre-trained Inception V3 model with transfer learning performed the best. In this work, VGG-16 model with Adam optimizer achieved the highest average accuracy with 99% and high accuracy of 91% for multi-class classification using InceptionV3 model.