Denoising Autoencoders for the Detection of Patients Out of Distribution of Healthy Individuals

This paper uses a Denoising AutoEncoder (DAE) to detect patients who are out of the distribution of healthy individuals.A DAE is trained only on healthy individuals.In the testing phase, the trained DAE is used to reconstruct healthy and pathological individuals with masked features.An error of reconstruction is computed on the masked features of the two families and compared.This paper shows the difference between the distributions of the error of reconstruction.This difference leads to the use of this method for classification.