Deep multilevel contextual networks for biomedical image segmentation

Abstract

Biomedical image segmentation plays a significant role in medical disease diagnosis and biological interconnection interpretation, such as connectome analysis. However, biomedical images are usually of formidable size, which renders human annotation in a large-scale way impractical to fulfill the whole job. An alternative way for biomedical image segmentation is to utilize computerized methods for automatic image analysis. Deep learning has advanced the performance of biomedical image segmentation dramatically. However, the scale of biomedical structures varies significantly and aggregating multilevel contextual information should be harnessed in an explicit way. To address this difficult problem, we propose a deep contextual network here by leveraging multilevel contextual information from the deep hierarchical structure to achieve better segmentation performance. To further improve the robustness against the vanishing gradients and strengthen the capability of the backpropagation of gradient flow, auxiliary classifiers are incorporated in the architecture of our deep neural network. It will be shown that our method can effectively parse the semantic meaning from the images with the underlying neural network and accurately delineate the structural boundaries with the reference of low-level contextual cues. Experimental results on the two benchmark datasets, including 2012 ISBI segmentation challenge of neuronal structures and 2015 MICCAI nuclei segmentation challenge, demonstrate that the proposed method can outperform the state-of-the-art methods by a large margin with respect to different evaluation measures.