Automated Deep Learning–based Detection and Segmentation of Lung Tumors at CT Imaging

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Background Detection and segmentation of lung tumors on CT scans are critical for monitoring cancer progression, evaluating treatment responses, and planning radiation therapy; however, manual delineation is labor-intensive and subject to physician variability. Purpose To develop and evaluate an ensemble deep learning model for automating identification and segmentation of lung tumors on CT scans. Materials and Methods A retrospective study was conducted between July 2019 and November 2024 using a large dataset of CT simulation scans and clinical lung tumor segmentations from radiotherapy plans. This dataset was used to train a 3D U-Net-based, image-multiresolution ensemble model to detect and segment lung tumors on CT scans. Model performance was evaluated on internal and external test sets composed of CT simulation scans and lung tumor segmentations from two affiliated medical centers, including single primary and metastatic lung tumors. Performance metrics included sensitivity, specificity, false positive rate, and Dice similarity coefficient (DSC). Model-predicted tumor volumes were compared with physician-delineated volumes. Group comparisons were made with Wilcoxon signed-rank test or one-way ANOVA. P < 0.05 indicated statistical significance. Results The model, trained on 1,504 CT scans with clinical lung tumor segmentations, achieved 92% sensitivity (92/100) and 82% specificity (41/50) in detecting lung tumors on the combined 150-CT scan test set. For a subset of 100 CT scans with a single lung tumor each, the model achieved a median model-physician DSC of 0.77 (IQR: 0.65-0.83) and an interphysician DSC of 0.80 (IQR: 0.72-0.86). Segmentation time was shorter for the model than for physicians (mean 76.6 vs. 166.1-187.7 seconds; p<0.001). Conclusion Routinely collected radiotherapy data were useful for model training. The key strengths of the model include a 3D U-Net ensemble approach for balancing volumetric context with resolution, robust tumor detection and segmentation performance, and the ability to generalize to an external site.