Performance-based multi-classifier decision fusion for atlas-based segmentation of biomedical images

Combinations of multiple classifiers have been found to be consistently more accurate than a single classifier. The construction of multiple independent classifiers, however, is typically a non-trivial problem. In atlas-based segmentation, multiple classifiers arise naturally, for example, from using multiple atlases. This paper evaluates the application of performance-based decision fusion methods to multi-classifier atlas-based segmentation. In a leave-one-out study, each of 20 subjects is segmented using each of the remaining 19 as the atlas. The resulting 19 segmentations per subject are combined into a final segmentation using three different methods: 1) simple decision fusion using the sum rule; 2) using a binary classifier performance model; 3) using a multi-label classifier performance model. The accuracy of each combined segmentation is computed by comparing it to the manual ground truth segmentation. The two methods that incorporate classifier performance outperform sum rule fusion, with the multi-label model performing better than the binary model.