Abnormality detection using deep neural networks with robust quasi-norm autoencoding and semi-supervised learning

Abstract

Abnormality detection in biomedical images is a one-class classification problem, where methods learn a statistical model to characterize the inlier class using training data solely from the inlier class. Typical methods (i) need well-curated training data and (ii) have formulations that are unable to utilize expert feedback through (a small amount of) labeled outliers. In contrast, we propose a novel deep neural network framework that (i) is robust to corruption and outliers in the training data, which are inevitable in real-world deployment, and (ii) can leverage expert feedback through high-quality labeled data. We introduce an autoencoder formulation that (i) gives robustness through a non-convex loss and a heavy-tailed distribution model on the residuals and (ii) enables semi-supervised learning with labeled outliers. Results on three large medical datasets show that our method outperforms the state of the art in abnormality-detection accuracy.