Brain Tumor Classification Using a Pre-Trained Auxiliary Classifying Style-Based Generative Adversarial Network

Abstract

Computer Vision's applications and their use cases in the medical field have grown vastly in the past decade. The algorithms involved in these critical applications have helped doctors and surgeons perform procedures on patients more precisely with minimal side effects. However, obtaining medical data for developing large scale generalizable and intelligent algorithms is challenging in the real world as multiple socio-economic, administrative, and demographic factors impact it. Furthermore, training machine learning algorithms with a small amount of data can lead to less accuracy and performance bias, resulting in incorrect diagnosis and treatment, which can cause severe side effects or even casualties. Generative Adversarial Networks (GAN) have recently proven to be an effective data synthesis and augmentation technique for training deep learning-based image classifiers. This research proposes a novel approach that uses a Style-based Generative Adversarial Network for conditional synthesis and auxiliary classification of Brain Tumors by pre-training. The Discriminator of the pre-trained GAN is fine-tuned with extensive data augmentation techniques to improve the classification accuracy when the training data is small. The proposed method was validated with an open-source MRI dataset which consists of three types of tumors - Glioma, Meningioma, and Pituitary. The proposed system achieved 99.51% test accuracy, 99.52% precision score, and 99.50% recall score, significantly higher than other approaches. Since the framework can be made adaptive using transfer learning, this method also benefits new and small datasets of similar distributions.

Keywords

• generative adversarial network
• image
• magnetic resonance imaging
• medical images
• network