Overview of Lymph Node Segmentation on Medical Images
The lymph nodes are part of the reticuloendothelial system. They are a vital part of the immune system and are significant actors in many cancers, both hematologic and solid. Typically a human has several hundreds of lymph nodes in their body. Most of the tissues in the body, except for the brain, have specific patterns of lymphatic drainage, with lymphatic channels draining into specific lymph node chains. Lymph nodes may become enlarged because of immune responses to infection/inflammation or because of infiltration with cancer cells, although tumor- involved nodes may also be normal in size. Such lymph nodes can be observed on medical imaging including computed tomography (CT) or magnetic resonance imaging (MRI) images. In clinical imaging, it is important to detect lymph nodes whose diameters are > 5 mm. Quantitative evaluation of lymph nodes is important for diagnosis and subsequent staging after surgical, medical, and/or radiation therapy.
Most enlarged lymph nodes can be identified as elliptical-shaped structures on CT images. Figure 3.12 shows examples of lymph nodes observed on axial CT slices. On contrast-enhanced CT images, these regions can be observed as foci whose intensity values are higher than surrounding structures such as fat. In the mediastinal area, the existing areas of the lymph nodes are almost fixed.
Several studies have been conducted for automated detection of nodes on chest and abdominal CT images [27,77]. The basic framework of detection consists of (a) blob-like structure enhancement for lymph node candidate detection and (b) falsepositive reduction.
Several methods have been proposed for “initial candidate” selection: (a) the Hessian-based method , (b) the directional difference filter-based method , and (c) the Haar-like feature method with machine learning augmentation . The Hessian-based approach tries to detect lymph node candidates by the fact enlarged lymph nodes show a spherical or elliptical shape on CT images. The directional difference filter method uses a similar approach. The Haar-like feature- based approach directly computes features that enlarged lymph nodes have to extract lymph node candidates.
Fig. 3.12 Examples of lymph nodes depicted on axial CT slice images
A feature-based approach is generally used to avoid false positives. Several classifiers are constructed to discriminate false-positive regions from lymph node candidate regions. Most methods use a machine learning approach to classify truepositive and false-positive regions.
A lymph node atlas showing the usual distribution of nodes assists in the falsepositive reduction process. Feuerstein et al. (2012)  used a patient-specific mediastinal atlas to reduce false positives.