Deep learning PET epilepsy detection with a novel symmetric loss convolutional autoencoder

Positron Emission Tomography imaging (PET) with 18F-fluoro-2-deoxyglucose (FDG) is a well established method to image brain glucose metabolism in healthy and disease states with application areas in neuro-degenerative and seizure disorders. Epilepsy is one of the most common neurological disorders effecting approximately 1 - 2% of the population. Advances in deep learning have revolutionized quantitative analysis and interpretation of medical images. Increasing availability of large medical imaging databases and sophisticated methods of extracting their discriminative features allows the potential for a greater understanding of the alterations of medical images in varying disease states. Probing medical images latent state presentations allows compressed representations with descriptive attributes to be explored. The auto-encoder and its variants serve this purpose; wherein they naively attempt to learn the identity function to reconstruct the input. Further constraints or regularization on the autoencoder avoid over-fitting to the training data hence boosting the discovery process. An autoencoder trained on a normal healthy database has recently found application areas in anomaly detection of FDG uptake with the magnitude of the reconstruction error serving as a proxy for abnormal brain patterns. In this work we take advantage of the fact that FDG uptake in the healthy subject is usually homogeneous and symmetrical with left-right asymmetries in activity concentration present in patients with hypometabolic epileptogenic regions. We therefore construct a novel autoencoder for anomaly detection in FDG brain PET and utilize a regularizing symmetry term in the loss function. This has the effect of producing increased reconstruction error in the event of an anomaly. The autoencoder is trained on 120 normal volunteers and tested on 3 patients with diagnosed epilepsy, demonstrating an average increase in reconstruction error of 6%, hence greater discriminative power in anomaly identification