Classification of optical tomographic images of rheumatoid finger joints with support vector machines

Over the last years we have developed a sagittal laser optical tomographic (SLOT) imaging system for the diagnosis and monitoring of inflammatory processes in proximal interphalangeal (PIP) joint of patients with rheumatoid arthritis (RA). While cross sectional images of the distribution of optical properties can now be generated easily, clinical interpretation of these images remains a challenge. In first clinical studies involving 78 finger joints, we compared optical tomographs to ultrasound images and clinical analyses. Receiver-operator curves (ROC) were generated using various image parameters, such as minimum and maximum scattering or absorption coefficients. These studies resulted in specificities and sensitivities in the range of 0.7 to 0.76. Recently, we have trained support vector machines (SVMs) to classify images of healthy and diseased joints. By eliminating redundancy using feature selection, we are achieving sensitivities of 0.72 and specificities up to 1.0. Studies with larger patient groups are necessary to validate these findings; but these initial results support the expectation that SVMs and other machine learning techniques can considerably improve image interpretation analysis in optical tomography.