Unsupervised 3D Brain Anomaly Detection

Anomaly detection (AD) is the identification of data samples that do not fit a learned distribution. As such, AD systems can help physicians to determine presence, severity, and extension pathology. Deep generative models, such as Generative Adversarial Networks (GANs), be exploited capture anatomical variability. Consequently, any outlier (i.e., sample falling outside distribution) detected an abnormality in unsupervised fashion. By using this method, we only detect expected or known lesions, but even unveil previously unrecognized biomarkers. To best our knowledge, study exemplifies first approach efficiently handle volumetric 3D brain anomalies one single model. Our proposal high-detail 2D f-AnoGAN model obtained by combining state-of-the-art GAN with refinement training steps. In experiments non-contrast computed tomography images from traumatic injury (TBI) patients, detects localizes TBI abnormalities area under ROC curve $$\sim $$ 75 $$\%$$ . Moreover, test potential method for detecting other low quality images, preprocessing inaccuracies, artifacts, presence post-operative signs (such craniectomy shunt). The has rapidly labeling massive imaging datasets, well identifying new