Biomedical Image Analysis with Segmentation and Reconstruction

Abstract

The biomedical image segmentation plays an important role in cancer diagnosis. Cell segmentation and brain tumor are important subtopics of biomedical image segment- ation. Manual annotation of biomedical image is time-consuming and suffers from observer bias. We propose two cell segmentation methods based on convolutional neural network. The first cell segmentation method employs both generator and dis- criminator to obtain dual-contour enhanced predictions. The predictions are further processed by watershed algorithm. The post-processing technique is still required and not learning based which constrains its capability of being robust and general to other dataset. Based on the above disadvantage, we propose an end-to-end con- volutional neural network segmentation architecture. In terms of this architecture, semantic branch and instance branch are jointly trained with the shared backbone. When accurate segmentation architectures are available, designing an efficient seg- mentation architecture is in great need. We design a segmentation network with dilated depthwise separable convolution block as its fundamental component and evaluate its performance on brain tumor segmentation dataset. Curvilinear structure appears at both natural and artificial objects. For example, the vessel structure of human body and aerial road map are categorized as curvilinear structure. Reconstruction of neuronal axon and dendrites is significant to conduct morphometric analysis and understand working mechanism of human brain. Many methods were proposed to tackle with neuron reconstruction, but they fail to propose an effective solution of reconstructing large-scale neuron image. An adaptive tracing method is designed to trace large 3D volume block by block. In addition, graph neural network based vessel reconstruction method is applied to remove redundant tracing points of vessel reconstruction.