Learning-Boosted Label Fusion for Multi-atlas Auto-Segmentation

Abstract

Structure segmentation of patient CT images is an essential step for radiotherapy planning but very tedious if done manually. Atlas-based auto-segmentation (ABAS) methods have shown great promise for getting accurate segmentation results especially when multiple atlases are used. In this work, we aim to further improve the performance of ABAS by integrating it with learning-based segmentation techniques. In particular, the Random Forests (RF) supervised learning algorithm is applied to construct voxel-wise structure classifiers using both local and contextual image features. Training of the RF classifiers is specially tailored towards structure border regions where errors in ABAS segmentation typically occur. The trained classifiers are applied to re-estimate structure labels at “ambiguous” voxels where labels from different atlases do not fully agree. The classification result is combined with traditional label fusion to achieve improved accuracy. Experimental results on H&N images and ribcage segmentation show clear advantage of the proposed method, which offers consistent and significant improvements over the baseline method.