Whole Regional Lymph Node Area Delineation with Deep Learning Model for Total Marrow and Lymphoid Irradiation.

Abstract

Total body irradiation (TBI) has been performed for conditioning before hematopoietic stem cell transplantation. However, TBI can be related to diverse adverse events including radiation pneumonitis and cataract. Efforts to reduce these events include the total marrow irradiation (TMI) and total marrow and lymphoid irradiation (TMLI). Compared to TMI, TMLI requires more target delineations with lymph nodes which can be labor-intensive and time-consuming. However, with the TMI plans, the coverage to lymph node might be lower than TMLI and its clinical significance is unknown. In the current study, we aimed to develop a deep learning model for automatic delineation of whole regional lymph nodes area and assess the dose coverage of lymph nodes with TMI plans. Whole regional lymph nodes (cervical, axillary, mediastinal, para-aortic, common iliac, external iliac, internal iliac, obturator, presacral, inguinal lymph nodes) were manually contoured by 3 radiation oncologists in 26 patients having whole body computed tomography (CT) images. Twenty patients were designated as the training/validation set and 6 patients as the testing set, and model was developed using the 'nnUNET' framework. The trained model was evaluated with dice coefficient score (DCS), precision, and recall. In addition, dose coverage of the automatically or manually delineated lymph nodes in TMI plans was calculated. The mean value of DCS, precision, and recall of the trained model was 0.76, 0.81, and 0.74, respectively. Dose parameters for manually delineated lymph nodes in previously treated TMI plans showed the mean value of V100% (the percentage of volume receiving 100% of the prescribed dose), V95%, and V90% to be 46.50%, 62.12%, and 73.68%, respectively. The highest V90% was observed in presacral (93.61%), axillary (90.40%), obturator (88.78%), and internal iliac lymph nodes (84.67%). In contrast, the lowest V90% was identified in inguinal (47.95%), cervical (61.69%), and para-aortic (65.75%) and external iliac lymph nodes (68.97%). For automatically delineated lymph nodes, the mean value of V100%, V95%, and V90% of TMI plan was 38.35%, 55.06%, and 67.84%, respectively. The difference with dose coverage of lymph node between delineated manually and automatically was not statistically significant. Automatic delineation of lymph node using deep learning showed the potential to reduce the labor-intensive process of TMLI. When treated with TMI, the coverage of inguinal, cervical, para-aortic and external iliac lymph nodes was lower than expected.