Automatic coronary calcium scoring in multi-center lung cancer screening trial

Abstract

Machine learning plays an essential role in medical image analysis. Supervised machine learning methods typically assume that the training data and task data have an identical distribution. However, this assumption may not always hold. In medical imaging, distributions might change due to for example, variations in the image acquisition protocols or pathology shown. Acquiring a (large) new representative training set for every set of images with a different distribution is often very time consuming or practically impossible. Instead, it would be advantageous to reuse the existing training data to analyze new data having different distribution. Methods enabling such transfer of knowledge are called transfer learning methods. In this thesis, we propose a novel transfer learning method, inspired by TrAdaboost, that uses an iterative weighted nearest neighbor classifier to extract knowledge from one distribution (source) to analyze the data originating from another distribution (target). Our method first identifies parts of the source data useful for the analysis of the target data using a small set of the labeled target data. The classifier is subsequently trained using these selected source and target sets. The method was applied to automatic coronary calcium scoring with chest CT scans acquired in a multi-center trial. The source and the target data were represented with scans acquired in two different centers with different CT scanners. Performance of the proposed system was compared with the system described by Isgum et al using standard nearest neighbor classification. The results showed that transfer learning approach improved classification. The achieved sensitivity (detected coronary calcifications) was high, but further investigation might be needed to reduce false positive rate.