Unsupervised Feature Learning for Outlier Detection with Stacked Convolutional Autoencoders, Siamese Networks and Wasserstein Autoencoders: Application to Epilepsy Detection

In this study we tackle the problem of detecting subtle epilepsy lesions in multiparametric (T1w, FLAIR) MR images considered as normal during a visual examination by a neurologist (MRI negative). We cast this problem as an outlier detection problem and adapt the framework proposed in [1]. It consists in learning a oc-SVM model for each voxel in the brain volume. We generalize this approach by proposing unsupervised deep architectures as feature extracting mechanisms in order to learn representations characterizing healthy subjects. We hypothesize that such architectures may capture features that allow to distinguish pathological voxels from the normal cases used in the training. As such, we exploit and compare three architectures, a novel configuration of siamese networks, stacked convolutional autoencoders and Wasserstein autoencoders. The models are trained on 75 healthy subjects and validated on 21 patients (with 18 MRI negatives) with confirmed epilepsy lesions achieving the best sensitivity of 62%.