Multiple Modality Fusion for Glaucoma Diagnosis

Abstract

Computer-aided diagnosis for diagnosis / screening makes use of artificial intelligence based analytical methodologies to handle patient data. Ocular disease Glaucoma is the major irreversible cause of blindness. Past works focus on building systems on single modality (personal data or major image features) and achieve limited success. At this time, there isn’t an effective and standard screening practice, which leads to more than half of the glaucoma cases are undiagnosed, which prevents the early treatment of the disease and is a big burden to the patients and health management.

Overcoming the limitation of the performance of single modality based system, a multiple modality fusion based glaucoma diagnosis approach is introduced and discussed in the paper by integrating patient personal data, major ocular image features, and important genome SNPs features in one system. Multiple kernel learning is used to integrate the features from different modalities; different kernel functions correspond to different modalities of the integrated data and therefore are treated as different aspects of similarity. 2,258 cases from a Singapore population study are tested and evaluated the multiple modality fusion based glaucoma diagnosis approach. Receiver Operating Characteristic curves are plotted to compare the approach’s performance with individual classifiers based on patient personal data, images, and genome SNPs separately. Instead of using the cross-validation Leave- One-Out approach, which may prone to statistical overtraining, this paper separates the training and testing dataset and gives a convincing analysis of the performance of the approach. This new testing approach clearly shows that the multiple modality fusion based glaucoma diagnosis approach is able to achieve an area under curve value better than the individual personal data, image and genome information components respectively.