2105 - CT-Based Deep Learning Radiomics in Predicting Disease-Free Survival and Overall Survival for Esophageal Squamous Cell Carcinoma Undergoing Radiotherapy: A Multicenter Study

Purpose/Objective(s):

To develop and validate integrated models combining deep learning (DL), radiomics, and dosiomics features from CT and radiation dose distribution (RDD) images to predict disease-free survival (DFS) and overall survival (OS) in ESCC patients undergoing radiotherapy (RT).

Materials/Methods:

A retrospective cohort of 220 ESCC patients treated with RT was enrolled from two hospitals. The cohort was divided into training (n=132), internal validation (n=60), and external validation (n=28) sets. Mean DFS and OS were 16.2, 21.0 months and 16.2, 16.5 months for ESCC patients from hospital one and two, respectively. Radiomics, dosiomics, and DL features were extracted from CT and RDD images to calculate survival signatures for each patient. Nomogram was developed by integrating radiomics plus dosiomics (R+D) signatures and DL-based radiomics plus dosiomics (DL_R+D) signatures with clinical risk factors. Models were evaluated using the concordance index (C-index).

Results: The C-index values of DFS and OS were improved to 0.76, 0.73, 0.80; and 0.83, 0.77, 0.77 by combining R+D and DL_D+R signatures in the training, internal validation, and external validation cohorts. The combined R+D+DL_R+D nomogram with clinical factors achieved C-indices of 0.76 (training), 0.73 (internal validation), and 0.78 (external validation) for DFS prediction, and 0.84, 0.78, and 0.79 for OS prediction, respectively. Moderate generalizability was observed in external validation (C-index: 0.77–0.79), though limited by the small external cohort (n=28).

Conclusion:

This study underscores the potential of integrating DL with radiomics and dosiomics for ESCC prognosis. While results are promising, broader validation and improved methodological transparency are essential for clinical translation.