Prediction of head and neck cancer with Deep Active Learning

Abstract

Collecting medical data is a costly task because the volume of data is often very large. This is even more complex considering that professionals need a lot of time to give a diagnosis for all these data. Commonly used Deep Learning techniques are not always enough to address this problem, since the number of attributes for each sample can be too large. Therefore, there is a need to find a way to analyze this information using as few instances as possible. This work focuses on the use of Deep Active Learning to automatically label new samples. It uses the good classification capacity of Deep Learning models together with the sample study provided by Active Learning. Thus, the cost of labeling is reduced while seeking to maintain performance. Deep Active Learning splits the dataset into a labeled and an unlabeled set. The labeled set is used to initialize the Deep Learning model, which is used to extract features from the unlabeled pool. To test the effectiveness of Deep Active Learning, we used a dataset with Head and Neck cancer samples. Our results show that it is possible to work with high dimensional data thanks to this method. Using 6 instances to classify per iteration, we were able to obtain an accuracy of 83% to 100%.