Disentangled representation learning for semi-supervised classi cation of chest x-ray images

Abstract

Acquisition of medical dataset is a di cult and expensive process as it require expertise of doctors, radiologists or other medical professionals. Chest X-rays are very common medical imaging technique, hence there are large repository of such images, but same cannot be said of other medical images. However, success of deep learning architecture primarily depends on the availability of large labeled datasets. Hence, semi-supervised learning that uses a small labeled dataset to develop a model for a large unlabeled dataset can be e ective particularly in medical imaging. A drawback of common semi-supervised learning models like wrapper methods (eg. self-training) or randomization methods is that latent encoding learnt by a deep neural models is ignored and the label prediction is dependent on the partially labeled dataset obtained from a supervised learning method in the input-output space. This approach ignores the representation learning of the model, which is the goal of any machine learning algorithm, to obtain a true data distribution of the input data. The stochastic state of the hidden state can contribute signi cantly to the prediction since the distribution of latent variable plays a major role in approximating the data distribution. This research primarily focussed on extracting the stochastic feature of the hidden state of the input images for classi cation using limited labeled dataset. A total of 50000 chest x-ray images were used for rst creating a baseline model of basic convolutional neural net supervised classi cation that obtained an overall accuracy of 80.07%. With only 5000 labeled images, the classi cation accuracy reduced to 60.97%. Then, using a variational auto-encoder based semi-supervised model considering 5000 labeled and 45000 unlabeled images, the accuracy up to 72.9% was obtained which suggested that the use of unlabeled images could contribute to the better generalization of a machine learning model. The thesis also compared the generative ability of auto-encoder and variational auto-encoder before selecting the VAE as our representation model. Keywords: Semi-supervised classi cation, Representation Learning, Deep learning, Variational Auto-Encoder, Convolutional Neural Network, Latent embedding