Biomedical imaging informatics in ocular disease diagnosis
Date of Issue2015
School of Computer Engineering
Computer aided diagnosis (CAD) system allows cost effective and prompt disease diagnosis, which has both clinical and social significance. Current ocular CAD systems typically account for only one type of data, e.g. medical image which may yield suboptimal accuracy as the training data itself lack the complete aspects for decision making. A new challenge in CAD research is to integrate the distinct attributes of clinical research that are provided by different types of biomedical data. By combining heterogeneous data sources, a CAD system would integrate the complementary pieces of information and provide a more holistic appreciation of the multiple risk factors, thus improves disease detection accuracy. This PhD study aims to fill in the blank by proposing an innovative system AODI (Automatic Ocular Disease Diagnosis through Biomedical Imaging Informatics), which focuses on CAD for ocular diseases, aiming to boost the diagnosis accuracy through intelligently combining image, SNP (Single-Nucleotide Polymorphism) and clinical data. AODI enables a data-driven approach that takes advantage of ever-growing heterogeneous data sources and improves the performance when more data or additional information becomes available. We investigate the recent advancements in kernel learning and deploy multiple kernel learning (MKL) algorithms for AODI. We conduct experiments to predict major ocular diseases including glaucoma, age-related macula disease (AMD), and pathological myopia (PM), using heterogeneous data sets covering image, SNP and clinical data which are obtained from a holistic population study conducted in Singapore. We also perform comprehensive statistical analysis to validate the improvement in the accuracy of predictions and prove the effectiveness of the proposed framework. To our best knowledge, AODI is the first published work using MKL to integrate multiple kinds of information including image, SNP and clinical data for ocular disease screening/diagnosis. Using MKL, the resulting classifier optimizes the contribution from each sub-kernel through learning an adapted kernel function from each of the heterogeneous feature sets. Such a framework paves a holistic way for automatic and objective disease diagnosis and screening. Moreover, our work on feature selection for SNP data tackles the challenge of SNP selection by innovatively grouping SNPs into functional groups (genes, interacting proteins and biological pathways), and thus explores the biomedical knowledge by sparse learning. Finally, we innovatively incorporates classemes (pre-learned classifiers trained from individual informatics domains) into MKL, and further improves the performance of ocular disease detection.
DRNTU::Engineering::Computer science and engineering