Regional Manifold Learning for Disease Classification

SRI Authors: Kilian Pohl

Citation

Ye, D. H., Desjardins, B., Hamm, J., Litt, H., & Pohl, K. M. (2014). Regional manifold learning for disease classification. IEEE Transactions on Medical Imaging, 33(6), 1236-1247.

Abstract

While manifold learning from images itself has become widely used in medical image analysis, the accuracy of existing implementations suffers from viewing each image as a single data point. To address this issue, we parcellate images into regions and then separately learn the manifold for each region. We use the regional manifolds as low-dimensional descriptors of high-dimensional morphological image features, which are then fed into a classifier to identify regions affected by disease. We produce a single ensemble decision for each scan by the weighted combination of these regional classification results. Each weight is determined by the regional accuracy of detecting the disease. When applied to cardiac MRI of 50 normal controls and 50 patients with reconstructive surgery of Tetralogy of Fallot, our method achieves significantly better classification accuracy than approaches learning a single manifold across the entire image domain.

Index Terms: Abnormality detection, cardiac magnetic resonance imaging (MRI), manifold learning, morphological classification, tetralogy of Fallot (TOF)


Read more from SRI