Brainomaly: Unsupervised Neurologic Disease Detection Utilizing Unannotated T1-weighted Brain MR Images.

Area of Science:

• Medical imaging analysis
• Deep learning for anomaly detection
• Neuroscience and neuroimaging

Background:

• Acquiring large annotated medical datasets for deep learning is challenging and costly, especially for rare diseases.
• Unsupervised anomaly detection methods reduce annotation effort but often perform poorly on neuroimages.
• Existing methods fail to leverage unannotated mixed datasets (healthy and diseased) effectively for neurologic conditions.

Purpose of the Study:

• To develop an unsupervised method for neurologic disease detection tailored to neuroimages.
• To improve the performance of anomaly detection using unannotated mixed datasets.
• To introduce a novel metric for model selection in the absence of annotated data.

Main Methods:

• Proposed Brainomaly, a Generative Adversarial Network (GAN)-based image-to-image translation framework.
• Utilized unannotated datasets containing both healthy and diseased subjects.
• Introduced a pseudo-Area Under the Curve (AUC) metric for model selection during inference.

Main Results:

• Brainomaly significantly outperformed state-of-the-art unsupervised anomaly detection methods.
• Demonstrated superior performance in detecting Alzheimer's disease and headaches.
• Ablation studies confirmed the effectiveness of the proposed components.

Conclusions:

• Brainomaly offers a powerful solution for unsupervised neurologic disease detection, particularly when annotated data is scarce.
• The method's tailored image-to-image translation and use of mixed data improve detection accuracy.
• The pseudo-AUC metric aids in robust model selection for real-world applications.