2174 - A Pragmatic Approach to Developing a Lung Nodal Station Atlas as Baseline Data for an Auto-segmentation Algorithm

H. T. Wong¹, A. J. Hope¹, and G. Liu²; ¹Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Toronto, ON, Canada, ²Division of Medical Oncology and Hematology, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada

Purpose/Objective(s): Numerous existing atlases on lung nodal stations for lung cancer showcase ideal diagnostic images, which may not accurately represent the anatomy of real patients undergoing radiotherapy (RT) planning. To establish a consistent lung nodal station map for a deep learning auto-segmentation algorithm, we developed a practical nodal station delineation atlas based on real-world patient data. Materials/

Methods: We conducted a targeted literature review of published lymph node atlases for lung cancer. Nodal station definitions from these articles were tabulated, and key descriptive images were included when available. Further clarification from atlas authors was sought if necessary. The recommended nodal station volumes were then applied to anonymized RT patient datasets using treatment planning software. An iterative process was employed until uniform application of the recommendations to the dataset was achieved. Slight deviations from the recommendations were permitted to accommodate the minor differences between atlases and patients.
Results: From the targeted literature search, we obtained six lung lymph node atlases. Two were excluded: one due to outdated information and the other due to limited global use. Of the remaining four, two were from a diagnostic radiology perspective, one from radiation oncologists, and the last was a collaboration between thoracic surgeons and radiologists. We were able to delineate the following nodal stations on our data set: levels 1R, 1L, 2R, 2L, 3A, 3P, 4R, 4L, 5, 6, 7, 8, 9R, 9L, 10R, 10L, 11R, and 11L. We took and applied the nodal delineation definitions which were clear cut and practical to the variable anatomies in our data set. The average time spent delineating a patient was 6 hours.
Conclusion: By synthesizing existing lung nodal station atlases and addressing inter-atlas variability while considering patient-to-patient anatomical variation, we developed a pragmatic set of nodal station guidelines. These guidelines enable consistent delineation of real-world patient datasets, serving as a baseline model for auto-segmentation learning.