IASLC lymph node station segmentation for regional lymph node staging of lung cancer

Introduction: The regional lymph node map published by the International Association for the Study of Lung Cancer (IASLC) is an established guideline for the accurate and reproducible assessment of the extent of regional lymph node metastases, whose use in tandem with the TNM classification is crucial for treatment planning and prognostication. Nevertheless, employing the IASLC map for PET/CT reporting requires detailed guideline interpretation and three-dimensional spatial localization of the prescribed anatomical boundaries, which can be subject to inter-site and inter-reader variability. Notably, deep learning methods have shown high accuracy for delineating anatomical structures in medical images automatically. We developed an automated method for segmenting IASLC lymph node stations in CT images, and validated it in an extended cohort for the purpose of regional lymph node staging with PET/CT.

Methods: Two experts in radiology and radiation oncology reviewed published IASLC guidelines and reached a consensus for the delineation of IASLC lymph node stations in CT images. Scans of 201 subjects with lung cancer, of which 101 acquired for radiotherapy planning and 100 acquired for diagnosis with ¹⁸F-FDG PET/CT, were reviewed by expert physicians who segmented IASLC stations 1 to 10. These were used as reference to develop an image segmentation neural network, with cases split between training (136), tuning (34), and hold-out testing (31). A separate cohort of 400 subjects who underwent ¹⁸F-FDG PET/CT imaging for the staging of lung cancer was used for independent validation. In this cohort, an expert nuclear medicine physician identified and segmented sites suspicious for cancer. The expert then assigned these a TNM classification (8^th edition) and, for regional lymph nodes, an IASLC station classification. The image segmentation neural network was applied to the latter CT images to segment lymph node stations and determine fully automatically the IASLC station classification of expert-identified regional nodes. The anatomical classification determined by network and the one assigned by the expert were then compared to assess the accuracy of regional lymph node staging with respect to station, zone, and N-class granularities at the lymph node level and with respect to the N-class at a subject level.

Results: The image segmentation network was able to delineate IASLC stations in close correspondence with expert delineations (by-station-mean average surface distance: 2.3mm for the hold-out testing dataset). Of 400 lung cancer patients in the validation cohort, 183 had complete information annotated by the expert and at least one positive regional lymph node (N1: 24 subjects, N2: 79, N3: 80), resulting in 925 total regional lymph nodes evaluated (N1: 175 nodes, N2: 459, N3: 291). At the lymph node level, overall classification accuracy was 67% for IASLC station granularity, 76% for zone granularity and 88% for N-class granularity. At the subject level (N-positive subjects), overall N stage classification accuracy was 94%.

Conclusions: An image segmentation network trained to segment IASLC lymph node stations in CT scans was able to automatically classify the anatomical localization and N-class of regional lymph nodes with good to excellent agreement with an expert physician. The proposed method is promising for supporting reproducible and accurate lung cancer staging.

• Download figure
• Open in new tab
• Download powerpoint