Improved detection of synovial boundaries in ultrasound examination by using a cascade of active-contours

Rheumatoid arthritis (RA) is a chronic multisystemic autoimmune disease, with an unclear etiopathogenesis. Its early diagnosis and activity assessment are essential to adjust the proper therapy. Among the different imaging techniques, ultrasonography (US) allows direct visualization of early inflammatory joint changes as synovitis, being also rapidly performed and easily accepted by patients. We propose an algorithm to semi-automatically detect synovial boundaries on US images, requiring minimal user interaction. In order to identify the synovia-bone and the synovia-soft tissues interfaces, and to tackle the morphological variability of diseased joints, a cascade of two different active contours is developed, whose composition corresponds to the whole synovial boundary. The algorithm was tested on US images acquired from proximal interphalangeal (PIP) and metacarpophalangeal (MCP) finger joints of 34 subjects. The results have been compared with a consensus manual segmentation. We obtained an overall mean sensitivity of 85±13%, and a mean Dice's similarity index of 80±8%, with a mean Hausdorff distance from the manual segmentation of 28±10 pixels (approximately 1.4±0.5mm), that are a better performance than those obtained by the raters with respect to the consensus.