Labeled data are necessary for supervised neural network (NN) training. However, supervised learning does not scale favorably, because human intervention for labeling large datasets is expensive. Here, we propose a method that introduces interventions on the training set, and enables NNs to learn features in a self-supervised learning (SSL) setting. The method intervenes in the training data by randomly changing image contrast and removing input image patches, thus creating a significantly augmented training dataset. This is fed into an autoencoder (AE) network, which learns how to reconstruct input images given variable contrast and missing patches of pixels. The proposed technique enables few-shot learning of most relevant image features by forcing NNs to exploit context information in a generative model. Here, we focus on a medical imaging application, where large labeled datasets are usually not available. We evaluate our proposed algorithm for anomaly detection on a small dataset with only 23 training and 35 test images of T2-weighted brain MRI scans from healthy controls (training) and tumor patients (test). We find that the image reconstruction error for healthy controls is significantly lower than for tumor patients (Mann-Whitney U-test, p < 10-10), which can be exploited for anomaly detection of pathologic brain regions by human expert analysis of reconstructed images. Interestingly, this still holds for conventional AE training without SSL, although reconstruction error distributions for healthy/diseased subjects appear to be less dissimilar (p<10-7). We conclude that the proposed SSL method may be useful for anomaly detection in medical imaging, thus potentially enhancing radiologists' productivity.
|