GAN augmentation for multiclass image classification using hemorrhage detection as a case-study.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Deep learning models require large, diverse datasets for effective training.
• Traditional data augmentation methods may be insufficient for complex, imbalanced medical datasets.
• Generative Adversarial Networks (GANs) show potential for generating synthetic data.

Purpose of the Study:

• To evaluate GANs as an alternative to traditional data augmentation for medical image classification.
• To assess the efficacy of conditional DCGAN (cDCGAN) for augmenting hemorrhage detection datasets.

Main Methods:

• Designed and implemented a conditional DCGAN (cDCGAN) model.
• Trained multiple GAN models in parallel for online augmentation.
• Compared GAN-based augmentation with traditional methods on a hemorrhage detection dataset.

Main Results:

• cDCGAN augmentation demonstrated improved performance for detecting rare epidural hemorrhages.
• GAN-augmented models outperformed traditionally augmented models with identical classifier configurations.
• The study highlights the limitations of traditional augmentation for imbalanced datasets.

Conclusions:

• GANs provide a more effective data augmentation strategy for complex and imbalanced medical imaging datasets.
• Advanced augmentation techniques like GANs are necessary to overcome limitations of traditional methods.
• This approach can enhance deep learning model performance in challenging clinical scenarios.