158 - Longitudinal CT Feature-Based Model for Predicting Local Recurrence Free Survival in Esophageal Cancer Patients Treated with Definitive Chemoradiotherapy: A Multicenter Study

Z. Yang¹, J. Gong², H. Yue³, J. Lu⁴, W. Huang⁵, and L. N. Zhao²; ¹Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xian, Shanxi, China, ²Department of Radiation Oncology, Xijing Hospital, Air Force Medical University, Xian, China, ³School of Computer Science and Engineering,Central South University, Changsha, China, ⁴Department of Radiation Oncology, Sichuan Cancer Hospital and Institution, Sichuan Cancer Center, School of Medicine, University of Electronic Science and Technology of China, Radiation Oncology Key Laboratory of Sichuan Province, Chengdu, China, ⁵Shandong Cancer Hospital & Institute, Jinan, Shandong, China

Purpose/Objective(s): Accurate prediction of local recurrence-free survival (LRFS) remains a great challenge for esophageal squamous cell carcinoma (ESCC) patients receiving definitive chemoradiotherapy (dCRT). However, the integration of longitudinal data holds significant potential for improving predictive capabilities. This study aims to develop and validate a deep learning model that incorporates longitudinal CT scans to accurately predict LRFS in patients with ESCC following dCRT. Materials/

Methods: We retrospectively collected 321 ESCC patients who underwent dCRT at our hospital and randomly divided them into training set (224) and internal validation set (91). Additionally, 202 patients from three other cancer hospitals were divided into two external validation sets (111 and 91 patients). The patients computed tomography (CT) images before and after dCRT were utilized. LRFS-related intratumoral and peritumoral radiomic features were extracted and selected through LASSO-Cox. Subsequently, we developed a correlation-driven disentanglement survival network (CDDSN), which integrated a ResNet block, base transformer, and basic CNN to effectively fuse shared and unique deep features from longitudinal CT images. The selected radiomic and deep features were then integrated to establish the final CDDSN model, the performance of which was validated using external sets. Furthermore, we developed a multimodal nomogram using independent clinical factors and the CDDSN signature in the training set, which was subsequently tested in three validation sets.
Results: Compared to the models utilizing CT images acquired before or after dCRT, the CDDSN model based on longitudinal CT scans achieved the highest performance in LRFS predictive with the C-index of 0.796 (95% CI: 0.741-0.851), 0.738 (95% CI: 0.675-0.801), 0.712 (95% CI: 0.661-0.758) and 0.708 (95% CI: 0.657-0.748) in the training set, internal testing set and the two external validation sets, respectively. Moreover, the CDDSN model integrating both intratumoral and peritumoral features outperformed the model solely using intratumoral features in each dataset. Additionally, our CDDSN model demonstrated superior fusion and predictive power in LRFS compared to the early fusion and late fusion methods in the two external validation sets (C-index: 0.712 vs 0.690 vs 0.681; 0.708 vs 0.661 vs 0.684), respectively. Furthermore, multivariable Cox regression analysis revealed that age, T stage, and absolute lymphocyte count were significantly correlated with LRFS (P<0.05). The multimodal nomogram, incorporating clinical factors and the CDDSN signature, exhibited a favorable predictive power for LRFS with a C-index of 0.826 in the training set and 0.709-0.734 in the validation sets.
Conclusion: The CDDSN model, utilizing longitudinal CT images, can enhance the predictive power for LRFS in patients with ESCC following dCRT. The CDDSN-based nomogram may offer a promising approach for helping ESCC individualized treatment decisions.