Augmenting healthy brain magnetic resonance images using generative adversarial networks

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Artificial intelligence in radiology

Background:

• Medical machine learning is hindered by data privacy concerns, leading to insufficient datasets.
• Brain tumor classification using medical images faces class imbalance issues, biasing models.
• Publicly available brain MRI datasets often lack sufficient 'no tumor' examples.

Purpose of the Study:

• To address the class imbalance problem in brain MRI datasets, specifically for the 'no tumor' class.
• To evaluate the effectiveness of Generative Adversarial Network (GAN)-based augmentation techniques against traditional methods.
• To compare classification performance using VGG16 on original, augmented, and combined datasets.

Main Methods:

• Employed Generative Adversarial Network (GAN) variants, including Deep Convolutional GAN (DCGAN) and Single GAN (SinGAN), for data augmentation.
• Implemented traditional rotation-based augmentation techniques for comparative analysis.
• Conducted VGG16 classification experiments on five distinct datasets: original, DCGAN-augmented, SinGAN-augmented, combined GAN-augmented, and rotation-augmented.

Main Results:

• The original, non-augmented dataset achieved the highest classification accuracy at 73%.
• Single GAN (SinGAN) demonstrated superior performance over Deep Convolutional GAN (DCGAN), with a 4% accuracy improvement.
• The non-augmented dataset exhibited the highest classification loss, highlighting the negative impact of class imbalance.

Conclusions:

• While GAN-based augmentation was explored, the original dataset performed best in this specific brain MRI classification task.
• SinGAN showed potential as an effective augmentation strategy, outperforming DCGAN.
• The study underscores the challenges of class imbalance in medical imaging and provides insights into augmentation technique efficacy.