MADGAN: unsupervised medical anomaly detection GAN using multiple adjacent brain MRI slice reconstruction

Abstract Background Unsupervised learning can discover various unseen abnormalities, relying on large-scale unannotated medical images of healthy subjects. Towards this, unsupervised methods reconstruct a 2D/3D single image to detect outliers either in the learned feature space or from high reconstruction loss. However, without considering continuity between multiple adjacent slices, they cannot directly discriminate diseases composed accumulation subtle anatomical anomalies, such as Alzheimer’s disease (AD). Moreover, no study has shown how anomaly detection is associated with stages, (i.e., more than two types of) diseases, multi-sequence magnetic resonance imaging (MRI) scans. Results We propose generative adversarial network (MADGAN), novel two-step method using GAN-based brain MRI slice anomalies at different stages structural MRI: ( Reconstruction ) Wasserstein loss Gradient Penalty + 100 $$\ell _1$$ ?1 loss—trained 3 axial slices next ones—reconstructs healthy/abnormal scans; Diagnosis Average _2$$ xmlns:mml="http://www.w3.org/1998/Math/MathML">?2 per scan discriminates them, comparing ground truth/reconstructed slices. For training, we use datasets 1133 T1-weighted (T1) and 135 contrast-enhanced T1 (T1c) scans for detecting AD metastases/various respectively. Our self-attention MADGAN very early stage, mild cognitive impairment (MCI), area under curve (AUC) 0.727, late stage AUC 0.894, while metastases T1c 0.921. Conclusions Similar physicians’ way performing diagnosis, massive training data, our first approach, MADGAN, reliably predict previous ones only images. As hyper-intense enhancing lesions, (especially late-stage)