Interactively learning a patient specific k-nearest neighbor classifier based on confidence weighted samples

An automatic segmentation method that fails for one scan of a patient is likely to fail in all follow up scans as well. We propose to construct a patient specific k-nearest neighbor classifier that learns from the test data while the user is interactively correcting the segmentation in the baseline scan. We apply the system to lung segmentation in chest CT scans. The system is set up in such a way that interaction is limited to single clicks in misclassified areas. Voxels indicated by a user as erroneously labeled are added to the training data. In classification, patient specific confidence weights are applied relative to the similarity between the test and training samples. The method is quantitatively validated on baseline and follow up scans from 16 patients. The results improve substantially in both baseline and follow up scans with only five clicks from the user in the baseline scan on average.