Abstract
BACKGROUND: Accurate segmentation of lung cancer gross tumor volume (GTV) on computed tomography (CT) is critical for radiotherapy planning yet remains difficult due to low tumor-tissue contrast, small target size, and high intratumoral heterogeneity. This study aimed to develop and validate an automatic method-a GTV segment-anything model generative adversarial network (GTV-SAMGAN)-for accurate, robust, and clinically efficient GTV segmentation on CT, with particular emphasis on small, low-contrast, and heterogeneous lesions. METHODS: We propose GTV-SAMGAN, built upon SAM medical 2D image (SAM-Med2D), integrating a newly developed GTV-minimal feature integration technology MFIT (GTV-MFIT) module with a GAN-based training scheme. The performance of GTV-SAMGAN was evaluated on a local clinical dataset and a public non-small cell lung cancer-radiomics (NSCLC-Radiomics) dataset (https://wiki.cancerimagingarchive.net/display/Public/NSCLC-Radiomics). We compared the proposed model against representative baselines (including SAM-Med2D and SwinU-Net) using the Dice coefficient, sensitivity, and specificity. RESULTS: On the local dataset, GTV-SAMGAN achieved a Dice coefficient of 83.74%, a sensitivity of 84.28%, and a specificity 99.98%, outperforming the other models. Compared to SwinU-Net, GTV-SAMGAN increased the Dice coefficient and sensitivity by 10.71% and 10.15%, respectively; compared to SAM-Med2D, it increased the Dice coefficient and sensitivity by 7.69% and 7.75%, respectively. On the NSCLC-Radiomics dataset, GTV-SAMGAN achieved a Dice coefficient of 82.92% and a sensitivity of 82.25%, representing an improvement over SAM-Med2D of 6.68% and 9.61%, respectively. CONCLUSIONS: By coupling SAM-Med2D with GTV-MFIT and GAN training, GTV-SAMGAN substantially improves lung cancer GTV segmentation, particularly for small and heterogeneous tumors, thereby enhancing the precision and efficiency of radiotherapy planning.