A Learning-Based Approach to Evaluate Registration Success

摘要

Clinical trials are more and more relying on medical imaging technologies to quantify changes over time during longitudinal studies. This calls for having an unsupervised batch registration process. However, even good registration algorithms fail, whether that is because of a small capture range, local optima, or because the registration finds an optimum that is not meaningful since the input data contains different anatomical sites. We propose a new method to evaluate the success or failure of batch registrations, so that failed or suspicious registrations can be flagged and manually corrected. The evaluation is based on a support vector machine that evaluates features representing the "goodness" of the registration result. We devise the features to be the distance measured between optima produced by different similarity measures as well as optima resulting from registering subsections of the volumes. The features of 30 volume registrations have been labeled manually and used for the learning phase. Based on a test on unseen 67 volume pairs of varying anatomical sites, we are able to classify 90% of the registrations correctly.