Anatomy Guided Thoracic Lymph Node Station Delineation in CT Using Deep Learning Model.

PURPOSE/OBJECTIVE(S) Accurate thoracic lymph node station (LNS) contouring is an essential step for N staging and radiotherapy, especially for those with lung and esophageal cancers. Thoracic LNS contouring is often determined using a system of anatomical descriptions proposed by the International Association for the Study of Lung Cancer (IASLC). Given that manual delineation of LNS can be prohibitively time-consuming and lead to large interobserver variability, an automated and accurate LNS delineation method is clinically useful. However, only a handful of computerized approaches existed with limited performance. Hence, we aim to develop an automated deep learning segmentation and parsing model by incorporating the key anatomic information to accurately delineate the LNS in CT. MATERIALS/METHODS We curated a dataset of contrast-enhanced CT scans with 12 manually annotated LNSs, ranging from station 1 to 8, from 97 esophageal cancer patients. Based on the IASLC protocol, the following key chest organs are also delineated: the arteries arch, ascendens and descendens, pulmonary artery, veins of inferior vena cava and superior vena cava, thyroid, trachea, bronchus, heart, lung, spine, and sternum. First, we train a nnUNet based deep model to segment the reference organs. Second, along the superior-inferior direction, we emulate radiation oncologist and use the first slice of the predicted sternum, the artery arch, and the artery pulmonary to divide the CT images into four sections. Third, we train another integrated segmentation model that combines the RTCT image, the segmented organ masks, and the four sections to delineate the LNS. The key organs provide the anatomical context in RTCT, while the four sections give references on the spatial locations of lymph node stations or zones. Four-fold cross-validation is used to fully evaluate the proposed deep network model. RESULTS The proposed framework has achieved 0.81 ± 0.06 Dice score (DSC), 0.9 ± 0.6 mm average surface distance (ASD) and 10.0 ± 3.6 mm Hausdorff distance (HD) on LNS parsing. Detailed results are shown in the Table. Our anatomy-guided method significantly outperforms the deep segmentation model using only RTCT images by a large margin of 4.8% in DSC and 19.0 mm in HD. CONCLUSION Our work demonstrates a well-calibrated deep learning framework that emulates radiation oncologists in LNS parsing with achieved good accuracy. We provide significant and tangible performance improvements, serving as a critical step towards deploying an automated, reliable and reproducible LNS parsing module into radiotherapy planning workflow.