Applying Generative Adversarial Network on Structural Brain MRI for Unsupervised Classification of Headache
Md Mahfuzur Rahman Siddiquee1, Jay Shah1, Todd Schwedt2, Catherine Chong2, Baoxin Li1, Teresa Wu1
1Arizona State University, 2Mayo Clinic
Objective:
This study aims to employ Brainomaly, an unsupervised detection approach, for classifying headache patients using unannotated T1-weighted brain MRIs. We introduce a pseudo-AUC metric for inference model selection without the need for annotated data.
Background:
Existing methods depend on annotated data for accurate headache detection. As labeled data are scarce, we propose Brainomaly, a Generative Adversarial Network-based approach that identifies structural abnormalities indicative of headaches, even without labeled data.
Design/Methods:
Brainomaly employs a generator network trained on two sets of brain MRIs: a combination of unannotated MRIs from headache patients and healthy controls and another set comprising solely MRIs from healthy controls. Through adversarial feedback from a discriminator network, the generator learns to represent 'healthy' brains and generate high-quality MRIs resembling healthy brains. The generated MRIs are subtracted from input MRIs to reveal structural deviations, aiding headache detection. We calculate the average value of the difference map as a headache detection score, where higher values indicate a higher likelihood of the MRI originating from a headache patient. To address model selection for inference, we propose the pseudo-AUC (AUCp) metric, assuming annotations for "unannotated mixed MRIs" as "headache," while known annotations are used for "healthy MRIs."
Results:
The dataset comprises 96 migraine, 48 acute post-traumatic headache, and 49 persistent post-traumatic headache patients, along with 532 healthy controls. Average age (headache 39.9+/-11.8 years, HC 41.6+/-12.7 years, P=0.1) and sex (P=0.1) did not differ between groups. Brainomaly achieved an average headache detection AUC (true) of 0.8960, surpassing existing unsupervised methods. Applying our AUCp metric to the second-best method, HealthyGAN, improved its AUC from 0.7695 to 0.8088. State-of-the-art methods such as ALAD, Ganomaly, and DDAD achieved lower AUCs of 0.6955, 0.6913, and 0.6280, respectively.
Conclusions:
These results prove Brainomaly's superiority in headache detection over existing methods. The proposed AUCp metric can select a better model for inference.