Contra-Discriminative GAN-based Anomaly Detection Framework for Medical Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated anomaly detection is crucial for medical applications like disease screening and quality control.
• Label-efficient methods are needed due to the scarcity of expert-labeled anomalous samples.
• Generative Adversarial Networks (GANs) are often used for anomaly detection by reconstructing normal images, but generated outputs may not perfectly match normal data distributions.

Purpose of the Study:

• To develop a novel Generative Adversarial Network (GAN) for improved anomaly detection in medical images.
• To address the limitations of existing GAN-based methods where generated normal images may not accurately represent the true normal data distribution.
• To leverage unlabeled data to enhance the performance of anomaly detection systems.

Main Methods:

• Introduction of a Contra-Discriminative Generative Adversarial Network (CD-GAN).
• Integration of a novel contrastive learning module within the GAN framework to align generated image distributions with normal image distributions.
• Training and evaluation using unlabeled images to guide the generation process.

Main Results:

• CD-GAN significantly outperforms state-of-the-art anomaly detection methods.
• Demonstrated superior performance across four public and one real-world clinical medical image datasets.
• The contrastive learning module effectively guides the generation process to better represent normal image distributions.

Conclusions:

• CD-GAN offers a powerful approach for label-efficient anomaly detection in medical imaging.
• The method shows competitive performance for diverse medical image anomaly detection tasks.
• This approach enhances the ability to discriminate between normal and anomalous samples using unlabeled data.