Development of a deep learning model based on computed tomography automatic segmentation to assist in selecting optimal time-to-surgery and dissected lymph node count for non-small cell lung cancer patients undergoing neoadjuvant immunotherapy and chemotherapy: a multicenter study.

Background: Many studies have confirmed the efficacy of neoadjuvant immunotherapy combined with chemotherapy (NICT) in treating patients with non-small cell lung cancer (NSCLC). However, the optimal time-to-surgery (TTS) and the appropriate dissected lymph node (DLN) count remain unclear. Therefore, this study aims to determine the optimal TTS and DLN count for NSCLC following NICT by establishing a deep learning model based on computed tomography (CT) automatic segmentation to predict the efficacy of neoadjuvant therapy. Methods: We retrospectively analyzed patients with NSCLC who underwent NICT and surgical treatment at two centers between January 2019 and June 2024. A proven high-precision and strong generalization three-dimensional (3D) segmentation architecture with a flexible interaction mode (VISTA3D) is applied in this study for CT images to achieve automatic tumor identification in NSCLC patients via automated segmentation combined with segmentation point prompts. This study employed ResNet18 (an 18-layer residual network pre-trained on ImageNet) as the feature extraction backbone. In the final configuration, the shallow stagelayers of the model were frozen, with only the two deeper stagelayers unfrozen to participate in gradient updates. This study employed BCEWithLogitsLoss to accommodate the classification requirements of the task. The primary evaluation metric in this study was the receiver operating characteristic curve. Based on the model, patients were divided into two groups: responders and non-responders. Optimal TTS and DLN counts were determined for each group. Results: A total of 330 patients were included, with 270 patients in the training set and 60 patients in the external test set. The area under the curve of the deep learning model was 0.854 (95% confidence interval: 0.745-0.976). We assigned patients to responders and non-responders groups based on the deep learning model score. In the responder group, prolonged TTS was associated with better prognosis, with no significant difference in postoperative complications. In the non-responder group, earlier surgery was associated with better prognosis, with no significant difference in postoperative complications. In the responder group, a DLN count of ≤21 was associated with better prognosis. Conclusions: In patients undergoing NICT, our model-based stratification suggests that for predicted responders, appropriately delaying surgery may be associated with better outcomes, whereas for predicted non-responders, earlier surgery is associated with better outcomes. Furthermore, among predicted responders, an observed DLN count of ≤21 was associated with improved survival, highlighting the potential importance of balancing oncologic resection with immune preservation. These findings require prospective validation.