Improving Multi-atlas Segmentation by Convolutional Neural Network Based Patch Error Estimation

Abstract

Multi-atlas segmentation (MAS) is widely used in automatically labeling medical images. The performance of patch-based MAS approaches relies on accurate estimation of local patch similarity, a proxy of the probability that an atlas patch provides the same label as the target patch. Learning-based image patch embedding techniques were recently proposed to transform raw intensity to feature maps and yield promising improvements compared to traditional raw intensity or hand-crafted features. In this study, we present a different approach in which the probability of atlas patch generating an erroneous vote, i.e. having a different label from the target patch, is directly estimated from the patches using a convolutional neural network (CNN). Experiments demonstrate that CNN-based estimates improve the segmentation accuracy of popular patch-based MAS techniques, i.e. spatially varying weighted voting and joint label fusion, in the context of segmenting medial temporal lobe subregions in T1-weighted MRI.