Main Session
Sep
29
PQA 07 - Head and Neck Cancer, Lung Cancer/Thoracic Malignancies, and Nursing and Supportive Care
3480 - Robust Automated Segmentation of Lung Tumors across Heterogeneous CT Protocols Using Deep Learning
Presenter(s)
William Green, MD - University of Pennsylvania Perelman School of Medicine, Radnor, Pa
W. R. Green1, R. McBeth1, E. Berlin1, V. Bourbonne2, S. Romdhani3, T. Theodoridis3, M. Costea4, G. Temiz3, A. Chakrabarti3, N. Paragios3, and E. Ozyar5; 1Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA, 2LaTIM, INSERM, UMR 1101, University of Brest, ISBAM, UBO, UBL, Brest, France, 3TheraPanacea, Paris, France, 4TheraPanacea, Lyon, France, 5Department of Radiation Oncology, Acibadem MAA University School of Medicine, Istanbul, Turkey
Purpose/Objective(s):
Manual delineation of lung tumors on CT images remains labor-intensive bottleneck in radiotherapy planning. This study evaluates a deep learning framework to provide fast and accurate automatic segmentation of lung lesions to enhance clinical efficiency.Materials/Methods:
A cascaded deep learning architecture, comprising a detector network followed by a 3D U-Net segmentation network [1] was trained on multi-institutional cohort comprising of 1,230 patients across eight independent datasets. The model was validated on a test set of 65 cases including contrast enhanced and non–contrasted CT images. Segmentation performance was evaluated using DICE similarity coefficient, 95th percentile Hausdorff distance (HD95), precision, and recall. To evaluate clinical acceptability, two expert radiation oncologists independently reviewed 30 cases using a three-point qualitative scale: (A) clinically acceptable, (B) acceptable with minor revisions, (C) clinically unacceptable.Results:
Overall mean DICE was 0.73 ± 0.19 with 0.80 precision and 0.74 recall. Elevated HD95 values were primarily attributable to false-positive segmentations, resulting in a mean HD95 of 43.4 ± 68mm. However, the median HD95 was 9.8 mm, supporting robust accuracy. Mean DICE was 0.71 ± 0.19 for contrast-enhanced images and 0.76 ± 0.19 for non–contrast images. In qualitative evaluation, 90% of cases were deemed clinically acceptable with no or with minor corrections. Only 3/30 cases were graded as C by both experts primarily due to lesion undercontouring. Interobserver agreement between the two oncologists was 77% with identical grades for 23 cases and a mix of A and B in 7 cases, indicating robust performance.Conclusion:
Our findings demonstrate that the AI-based framework provides accurate and generalizable lung lesion segmentation regardless of contrast usage across heterogeneous datasets and imaging protocols. High quantitative accuracy and strong clinical acceptability support its potential to reduce contouring burden, minimizing interobserver variability, and improve efficiency in radiotherapy planning. Reference: [1] Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 5–9 October 2015; pp. 234–241