In biomedical informatics, a large amount of image data has been collected to support clinical diagnosis, treatment decision, and medical prognosis. The large volume and the diversity of informatics across different imaging modalities require advanced and high-throughput image computing technologies to provide more accurate disease detection, deeper understanding of the mechanisms of disease progression, and better healthcare in precision medicine. With the ever-increasing amount of biomedical image data, it is very critical to design and develop efficient technologies for large-scale biomedical image data analysis.

In this dissertation, we will describe a high-throughput biomedical image computing framework for digital health, focusing on two important topics: object detection and segmentation as well as their applications, image understanding, in medical diagnosis. Specifically, given a pathology image, it begins with an efficient algorithm to detect individual objects (e.g., nuclei or cells), which will be used to conduct contour or shape initializations. Next, a novel segmentation algorithm is exploited to separate individual object by combining a scalable shape prior model and a fast repulsive deformable model. The shape prior modeling is formulated as a sparsity-based dictionary learning problem, which can be solved with efficient sparse encoding algorithms. The fast deformable model is obtained by incorporating a scalable sparse manifold learning algorithm into a local active contour model. In order to address the problem of limited training data in the medical domain, we present a transfer learning based dictionary learning for shape modeling, which can be applicable to multiple microscopy image datasets. We have compared the proposed approaches with many recent state of the arts and the experimental results demonstrate their effectiveness and efficiency.