f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Expert annotation of clinical images is time-consuming and may miss unknown markers.
• Supervised learning is limited by the scope of annotated lesions and their visual variability.

Purpose of the Study:

• To develop an unsupervised learning approach for identifying anomalous images and segments in clinical imaging.
• To propose fast AnoGAN (f-AnoGAN) as a generative adversarial network (GAN)-based method for anomaly detection and biomarker candidate identification.

Main Methods:

• Trained a generative model on healthy training data.
• Developed a fast mapping technique using a trained encoder to map new data to the GAN's latent space.
• Detected anomalies using a combined score including discriminator feature residual and image reconstruction errors.

Main Results:

• f-AnoGAN demonstrated high anomaly detection accuracy on optical coherence tomography (OCT) data, outperforming alternative methods.
• Generated images were indistinguishable from real normal retinal OCT images in a visual Turing test with experts.

Conclusions:

• f-AnoGAN offers an effective unsupervised approach for anomaly detection in clinical imaging.
• The method shows promise for identifying novel imaging biomarkers without prior knowledge of all possible findings.