A Multi-Armed Bandit to Smartly Select a Training Set from Big Medical Data

摘要

With the availability of big medical image data, the selection of an adequatetraining set is becoming more important to address the heterogeneity ofdifferent datasets. Simply including all the data does not only incur highprocessing costs but can even harm the prediction. We formulate the smart andefficient selection of a training dataset from big medical image data as amulti-armed bandit problem, solved by Thompson sampling. Our method assumesthat image features are not available at the time of the selection of thesamples, and therefore relies only on meta information associated with theimages. Our strategy simultaneously exploits data sources with high chances ofyielding useful samples and explores new data regions. For our evaluation, wefocus on the application of estimating the age from a brain MRI. Our results on7,250 subjects from 10 datasets show that our approach leads to higher accuracywhile only requiring a fraction of the training data.