Patient-Level Cross-validated nnU-Net for Multiclass Segmentation of Lung Parenchyma and Solid Adenocarcinoma on Thoracic CT.
Authors
Affiliations (4)
Affiliations (4)
- LPHE-Modeling and Simulation, Faculty of Sciences, Mohammed V University, Rabat, Morocco. [email protected].
- Sciences and Engineering of Biomedicals, Biophysics and Health Laboratory, Higher Institute of Health Sciences, Hassan First University, Settat, 26000, Morocco.
- Higher Institute of Nursing Professions and Health Techniques, Rabat, Morocco.
- LPHE-Modeling and Simulation, Faculty of Sciences, Mohammed V University, Rabat, Morocco.
Abstract
Automated segmentation of lung parenchyma and solid lung adenocarcinoma on thoracic computed tomography (CT) is needed for reproducible quantitative imaging, radiomics extraction, and treatment planning-related research. However, tumor segmentation remains challenging because of small lesion size, irregular morphology, partial-volume effects, and boundary ambiguity near vessels, pleura, atelectasis, and mediastinal structures. This study evaluated a single-input-channel multiclass nnU-Net framework for joint lung parenchyma and solid adenocarcinoma segmentation in pre-treatment CT examinations from 150 patients with pathologically confirmed solid lung adenocarcinoma. The thoracic CT volume was the only model input. Reference segmentation maps encoded three mutually exclusive classes: background (label 0), lung parenchyma (label 1), and tumor (label 2). The model was trained using the 3D full-resolution nnU-Net configuration and evaluated using patient-level fivefold cross-validation. Performance was quantified using the Dice similarity coefficient (DSC), intersection-over-union (IoU), 95th percentile Hausdorff distance (HD95), relative volume difference (RVD), and volume similarity (VS). Lung segmentation was highly accurate and stable, with median DSC > 0.97, IoU > 0.95, HD95 < 5 mm, and VS > 0.99 across folds. Tumor segmentation achieved moderate overlap performance, with median DSC values of approximately 0.70-0.78 and median IoU values of approximately 0.54-0.64, but substantial HD95 variability indicated residual boundary-localization failures in challenging cases. The single-channel multiclass framework is promising as a segmentation backbone for quantitative CT analysis; however, external validation, direct comparison with alternative models, and improved false-positive suppression are required before direct clinical use.