Deep learning model based on automated CT image segmentation for predicting the optimal radiotherapy protocol in limited-stage small cell lung cancer patients: a multicenter study.

Background: Radiotherapy combined with chemotherapy is the standard treatment regimen for patients with limited-stage small cell lung cancer (LS-SCLC), but the optimal radiotherapy regimen for these patients remains unclear. This study aims to construct an integrated, two-stage "image segmentation-response prediction" deep learning (DL) framework. By leveraging automated tumor segmentation to accurately define tumor habitats, this framework extracts deep imaging features to predict the optimal radiotherapy regimen for LS-SCLC patients. Methods: This study enrolled patients with LS-SCLC who received chemoradiotherapy at two centers. Center A (Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences) serves as the development cohort, while Center B (Xiaogan Central Hospital) functions as the external validation cohort. To establish the prediction framework, we first employed the VISTA3D model, integrating automatic and point-prompt guidance, to achieve highly precise three-dimensional (3D) tumor segmentation of the primary tumor lesion and its peritumoral habitats. Subsequently, the segmented target regions were fed into a ResNet18-based dual-branch DL model to extract deep radiomic features and generate treatment preference scores, thereby establishing a highly automated and interactive DL model to predict the optimal radiotherapy regimen by differentiating the expected progression-free survival (PFS) between hyperfractionated and conventionally fractionated regimens. Model performance is evaluated using metrics such as area under the curve (AUC). Results: This study involved 446 patients. When performing automatic segmentation using the VISTA3D segmentation model, the model achieved a Dice coefficient of 0.86±0.17 on the external validation cohort, with an AUC of 0.88 [95% confidence interval (CI): 0.71-0.96]. Among patients with a score >0.5 (indicating a predicted survival benefit from hyperfractionated radiotherapy), PFS and overall survival (OS) analyses demonstrated that patients receiving hyperfractionated radiotherapy showed significantly superior outcomes compared to those receiving conventionally fractionated radiotherapy [hazard ratio (HR) =0.55, P<0.001; HR =0.39, P<0.001]. In patients with a score ≤0.5 (indicating no predicted superior benefit from hyperfractionation), there was no significant difference in PFS and OS between those receiving hyperfractionated radiotherapy and those receiving conventional fractionated radiotherapy (HR =1.12, P=0.83; HR =1.14, P=0.85). However, the incidence of grades 3-4 acute esophagitis was significantly lower in patients receiving conventional fractionation radiotherapy (16.0% Conclusions: The DL model developed in this study, based on automatic segmentation combined with point-prompt technology, can replace prediction models based on manual segmentation. This model effectively assists clinicians in predicting the optimal radiotherapy regimen for LS-SCLC patients, validating the feasibility of a two-stage automated and interactive prediction framework integrating "image segmentation-response prediction". This approach can provide support for clinical decision-making in patients with LS-SCLC, advancing the process of personalized and precision cancer treatment.