Segmentation of lymph node in longitudinal studies of thoracic CT images

The assessment of lymph node is commonly used as a yardstick in diagnosis, staging, treatment and therapy control of tumours and their metastases. Current practice of the radiologists is to manually measure the major and minor axis of nodes based on experiences. This method is highly subjected to inter-operator variation. In this report, we present an automated lymph node segmentation model in longitudinal studies. With this model, lymph node would be segmented once using the semi-automated active contour (snake) at baseline and subsequent follow-up studies could be segmented automatically through registration. We applied two types of registration i.e. Free Form Deformation (FFD) and Demons Registration in this study and compared the efficiency and accuracy of both. It was found that Demons-based registration generates results with higher accuracy. We applied the automated segmentation technique to 18 pairs of lymph nodes from 8 patients at baseline and follow-up. A set of manual segmentation results, recognized by two experienced radiologists, serves as the gold standard for evaluation of performance of the method. In Demons-based model, we could observe that 11 out of 18 nodes are well segmented, 6 out of 18 nodes showed slight leakage due to blurred boundaries and 1 out of 18 failed terribly. Including the failed node in statistics, the overall automated segmentation showed an average Dice’s Coefficient of 0.799, sensitivity of 0.799 and specificity of 0.970.Bachelor of Engineering (Chemical and Biomolecular Engineering