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

Purpose: In recent years, the development and exploration of deeper and more complex deep learning models has been on the rise. However, the availability of large heterogeneous datasets to support efficient training of deep learning models is lacking. While linear image transformations for augmentation have been used traditionally, the recent development of generative adversarial networks (GANs) could theoretically allow us to generate an infinite amount of data from the real distribution to support deep learning model training. Recently, the Radiological Society of North America (RSNA) curated a multiclass hemorrhage detection challenge dataset that includes over 800,000 images for hemorrhage detection, but all high-performing models were trained using traditional data augmentation techniques. Given a wide variety of selections, the augmentation for image classification often follows a trial-and-error policy. Approach: We designed conditional DCGAN (cDCGAN) and in parallel trained multiple popular GAN models to use as online augmentations and compared them to traditional augmentation methods for the hemorrhage case study. Results: Our experimentations show that the super-minority, epidural hemorrhages with cDCGAN augmentation presented a minimum of 2x improvement in their performance against the traditionally augmented model using the same classifier configuration. Conclusion: This shows that for complex and imbalanced datasets, traditional data imbalancing solutions may not be sufficient and require more complex and diverse data augmentation methods such as GANs to solve. (C) 2022 Society of Photo-Optical Instrumentation Engineers (SPIE)