HealthyGAN: Learning from Unannotated Medical Images to Detect Anomalies Associated with Human Disease

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Automated anomaly detection in medical images (e.g., MRIs, X-rays) aids disease diagnosis.
• Current methods often rely solely on healthy subjects, which is insufficient when unannotated mixed datasets are available.
• Manual annotation by experts is costly and time-consuming, limiting the scalability of supervised methods.

Purpose of the Study:

• To enhance unsupervised anomaly detection in medical images.
• To leverage both healthy and unannotated mixed datasets for improved diagnostic models.
• To address the limitations of existing methods in real-world scenarios with abundant unannotated data.

Main Methods:

• Proposed HealthyGAN, a novel one-directional image-to-image translation technique.
• HealthyGAN learns to translate images from a mixed dataset (healthy and diseased) to resemble only healthy images.
• Anomaly detection is performed by generating a difference map between an input image and its translated healthy version.

Main Results:

• HealthyGAN significantly outperforms conventional state-of-the-art anomaly detection methods.
• Demonstrated superior performance on COVID-19, NIH ChestX-ray14, and Mayo Clinic datasets.
• The one-directional approach overcomes cycle-consistency limitations with unannotated data.

Conclusions:

• HealthyGAN offers an effective solution for unsupervised anomaly detection using readily available unannotated medical image data.
• The method enhances diagnostic capabilities by identifying anomalies without requiring expert annotations.
• This approach holds significant potential for improving the efficiency and accuracy of disease diagnosis in clinical practice.