Non-Rigid Registration And Robust Principal Component Analysis With Variation Priors: A High-Throughput Mouse Phenotyping Approach

Intensive global efforts are underway towards phenotyping the mouse genome, by systematically knocking out all approximate to 25,000 genes for comparative study. Analytical work of this scale easily overwhelms the traditional method using histological examination, leading to a significant demand for high-throughput approaches, especially via image informatics to efficiently identify phenotypes concerning morphological anomaly. We propose a high-throughput batch-wise anomaly detection framework without prior knowledge of the phenotype and the need for segmentation. Anomaly detection is centered on feature decomposition using robust principal component analysis (RPCA), which has previously been applied to many computer vision domains. However, baseline RPCA does not work well in the biomedical domain due to substantial natural variation in imaging data. In contrast, we develop a modified version (RPCA-P) that incorporates variation priors, coupled with non-rigid image registration to achieve this goal.