Enhanced MRI-based brain tumour classification with a novel Pix2pix generative adversarial network augmentation...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Medical imaging dataset scarcity and patient privacy concerns hinder AI-driven disease prediction.
• Existing tools can extract patient information from medical images, raising confidentiality issues.

Purpose of the Study:

• To propose generative adversarial networks (GANs) for synthetic data generation to overcome data limitations and privacy issues in brain tumour classification.
• To introduce and evaluate a novel Pix2Pix GAN model for generating synthetic brain tumour datasets.
• To develop and assess a conditional deep convolutional neural network (CNN) for classifying brain tumours using both synthetic and real datasets.

Main Methods:

• Utilized a novel Pix2Pix GAN for image-to-image translation to generate synthetic brain tumour datasets.
• Focused on classifying four types: glioma, meningioma, pituitary, and healthy brain tissues.
• Developed a novel conditional deep CNN architecture for processing and classifying both synthetic and original Kaggle datasets.

Main Results:

• The conditional deep CNN achieved 86% accuracy using synthetic images for brain tumour classification.
• Comparative analysis demonstrated superior performance of the proposed conditional deep CNN over state-of-the-art models (ResNet50, VGG16, VGG19, InceptionV3).
• The Pix2Pix GAN augmentation technique proved effective in creating high-quality synthetic datasets.

Conclusions:

• The Pix2Pix GAN-based synthetic data augmentation is effective for accurate brain tumour classification.
• The developed conditional deep CNN shows high performance in brain tumour detection, diagnosis, and classification.
• This approach offers a promising solution for improving AI-based disease prediction and treatment planning while ensuring patient confidentiality.