Image segmentation is a fundamental problem in biomedical image analysis. In recent years, deep learning methods have emerged and continued to prosper as powerful tools for natural scene image segmentation. While biomedical image segmentation is in close relation to natural scene image segmentation, general deep learning methods for natural scene images may not work well on biomedical applications because of two unique properties of biomedical images. First, biomedical images can be true 3D volumetric images, which post significant challenges on computational approaches and computation resources, while most natural scene images are 2D images. Second, only trained biomedical experts can annotate biomedical images well, which makes it quite expensive and difficult to obtain sufficient training data. In this dissertation, we develop new deep learning based approaches specifically targeting these two unique properties of biomedical images. To achieve accurate segmentation on 3D biomedical images, we propose a new ensemble learning framework to unify the merits of 2D and 3D deep learning models. To reduce the annotation burden for biomedical images, we propose novel methods to (1) improve the effectiveness of manual annotation by suggesting the most useful object samples to annotate and (2) improve the efficiency of manual annotation by allowing inexact "rough" labeling. Extensive experiments on various biomedical image segmentation datasets show that our methods can achieve state-of-art performance while reducing significant amounts of annotation efforts.