Multi-class Generative Adversarial Networks: Improving One-class Classification of Pneumonia Using Limited Labeled...

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging Analysis

Background:

• Generative Adversarial Networks (GANs) excel at image generation and class distribution learning.
• GANs are applied in semi-supervised medical image analysis, including detecting diseases like pneumonia.
• A key challenge arises when image classes share similarities, causing GANs to generalize and impair classification accuracy.

Purpose of the Study:

• To demonstrate how GAN generalization affects classification accuracy with visually similar datasets.
• To illustrate GAN-induced misclassification of pneumonia in X-rays as healthy cases.
• To propose a GAN training modification for enhanced classification of similar image classes within a semi-supervised framework.

Main Methods:

• Utilized MNIST and Fashion-MNIST datasets to visually inspect GAN generalization on similar images.
• Applied GANs to semi-supervised pneumonia detection in X-rays to observe misclassification.
• Developed and tested a modified GAN training strategy using limited labeled data.

Main Results:

• Confirmed that GANs generalize and lead to poor classification when faced with visually similar image classes.
• Observed misclassification of pneumonia X-rays as healthy cases due to GAN generalization.
• The proposed GAN modification demonstrated improved classification performance for similar image classes.

Conclusions:

• GANs' generalization capability poses a significant challenge for classifying visually similar medical images.
• The modified GAN training approach effectively enhances classification accuracy in semi-supervised learning for similar classes.
• This research offers a promising solution for more reliable medical image analysis using GANs.