CT-Based Deep Learning Model for Predicting Local Recurrence-Free Survival in Esophageal Squamous Cell Carcinoma Patients Received Concurrent Chemo-Radiotherapy: A Multicenter Study

Abstract

<h3>Purpose/Objective(s)</h3> For esophageal squamous cell carcinomas (ESCC) patients received concurrent chemo-radiotherapy (CCRT), local recurrence is the most common failure pattern and reliable markers for prognosis are lacking. Previous studies have demonstrated the predictive role of traditional radiomics features for prediction of local recurrence-free survival (LRFS) in ESCC. Nevertheless, traditional radiomics based on human-defined handcrafted features may not fully characterize tumor heterogeneity. Some studies have provided evidence that deep learning with advantages in voxel analysis could provide remarkable performance. This study aims to establish and validate a deep learning model for predicting LRFS in ESCC patients received CCRT. <h3>Materials/Methods</h3> We retrospectively included 302 patients from Xijing Hospital and randomly divided them into training set (201) and internal validation set (101) according to 2:1. 95 patients from Tianjin Cancer Hospital and Shandong Cancer Hospital were included as the external validation set. All patients underwent a contrast-enhanced computed tomography (CE-CT) scan before CCRT and were followed up for more than 24 months after CCRT. The deep learning model was developed by using 3D-Densenet deep learning architecture. Manually segmented tumors based on CE-CT were used as model input, LRFS was the prediction target. The deep learning signature was built using the deep-score output by the model. <h3>Results</h3> The median follow-up time of all patients was 26.77 months and 257 of 397 patients (64.74%) were confirmed local recurrence or death during the follow-up period. The deep learning model for prediction of LRFS in ESCC patients received CCRT showed good prognostic performance, with a C-index of 0.7337 (95% CI: 0.6800–0.7874) in the training set, which was validated in the internal (0.7203 [95% CI: 0.6450–0.7957]) and external validation (0.7167 [95% CI: 0.6416–0.7918]) sets, respectively. Kaplan-Meier survival analysis showed that the median of deep-score (-0.06) could stratify patients into high and low-risk groups for different LRFS. The low-risk group with the lower deep-score had a significantly higher LRFS than that of the high-risk group with a high deep-score (2-year LRFS 71.1% vs 33.0%, p<0.0001) in the training set. The result was validated in the internal (2-year LRFS 58.8% vs34.8%, p<0.01) and external validation (2-year LRFS 61.9% vs 22.4%, p<0.0001) sets, respectively. <h3>Conclusion</h3> Deep learning signature can be used as a non-invasive radiomics marker to predict LRFS in ESCC patients received CCRT. This is the first multicenter-based study using deep learning to predict local recurrence in esophageal cancer.