[Paper Review] Inception Augmentation Generative Adversarial Network

Saman Motamed, Farzad Khalvati|arXiv (Cornell University)|Jun 5, 2020

COVID-19 diagnosis using AI22 references4 citations

TL;DR

This paper proposes an Inception Augmentation Generative Adversarial Network (IAGAN) that generates realistic pneumonia-class chest X-rays by leveraging both normal and pneumonia images during training. By using a GAN to augment data for a specific anomaly detection task, the method improves AnoGAN's performance, increasing AUC from 0.83 to 0.88 in pneumonia detection on small datasets.

ABSTRACT

Successful training of convolutional neural networks (CNNs) requires a substantial amount of data. With small datasets, networks generalize poorly. Data Augmentation techniques improve the generalizability of neural networks by using existing training data more effectively. Standard data augmentation methods, however, produce limited plausible alternative data. Generative Adversarial Networks (GANs) have been utilized to generate new data and improve CNN performance. Nevertheless, generative models have not been used for augmenting data to improve the training of another generative model. In this work, we propose a new GAN architecture for semi-supervised augmentation of chest X-rays for the detection of pneumonia. We show that the proposed GAN can augment data for a specific class of images (pneumonia) using images from both classes (pneumonia and normal) in an image domain (chest X-rays). We demonstrate that using our proposed GAN-based data augmentation method significantly improves the performance of the state-of-the-art anomaly detection architecture, AnoGAN, in detecting pneumonia in chest X-rays, increasing AUC from 0.83 to 0.88.

Motivation & Objective

• To address poor generalization in CNNs trained on small medical imaging datasets, particularly for pneumonia detection in chest X-rays.
• To overcome the limitations of standard data augmentation, which produces only minor, non-plausible variations of existing data.
• To explore the use of generative models not just for data generation, but as a tool to augment data for training another generative model.
• To improve the performance of the state-of-the-art anomaly detection model AnoGAN through class-specific data augmentation using a novel GAN architecture.

Proposed method

• The proposed IAGAN architecture uses a conditional GAN framework that generates synthetic pneumonia-like chest X-rays by conditioning on both normal and pneumonia images.
• It incorporates an Inception module to enhance feature representation learning and improve the quality of generated images.
• The generator is trained to produce realistic, class-specific augmentations that preserve anatomical plausibility and disease patterns.
• The discriminator is trained to distinguish real from generated images while also being guided by the feature space of the target anomaly detection model (AnoGAN).
• The method employs a semi-supervised learning strategy, using both labeled normal and pneumonia images to train the GAN for class-specific augmentation.
• The generated data is used to fine-tune AnoGAN, improving its ability to detect anomalies in low-data regimes.

Experimental results

Research questions

• RQ1 Can a GAN be effectively used to generate realistic, class-specific data augmentations for a downstream anomaly detection model?
• RQ2 Does GAN-based data augmentation improve the performance of AnoGAN in detecting pneumonia on small chest X-ray datasets?
• RQ3 Can a GAN leverage both normal and pneumonia images to generate plausible synthetic pneumonia images for data augmentation?
• RQ4 How does the proposed IAGAN method compare to standard data augmentation techniques in terms of AUC improvement for pneumonia detection?

Key findings

• The proposed IAGAN-based data augmentation significantly improves AnoGAN's performance on pneumonia detection in chest X-rays.
• The AUC of AnoGAN increased from 0.83 to 0.88 when trained with IAGAN-generated data, demonstrating a measurable improvement in detection capability.
• The generated images were realistic and preserved disease-specific patterns, enabling effective generalization during training.
• The method outperformed standard data augmentation techniques by generating more diverse and plausible training samples.