Contextual convolutional neural network filtering improves EM image segmentation

Xundong Wu, Yong Wu, Ligia Toro, Enrico Stefani

Abstract: We designed a contextual filtering algorithm for improving the quality of image segmentation. The algorithm was applied on the task of building the Membrane Detection Probability Maps (MDPM) for segmenting electron microscopy (EM) images of brain tissues. To achieve this, we executed supervised training of a convolutional neural network to recover the ground-truth label of the masked-out center pixel from patches sampled from an un-refined MDPM. Through this training process the model learns the distribution of the segmentation ground-truth map . By applying this trained network over MDPMs we are able to integrate contextual information to obtain with better spatial consistency in the high-level representation space. By iteratively applying this network over the MDPMs for multiple rounds, we were able to significantly improve the EM image segmentation results.

Conflicts: ucla.edu, usc.edu