Multivariable Radiomics Model for Predicting Programmed Death-Ligand 1 Expression After Neoadjuvant Chemoradiotherapy in Esophageal Squamous Cell Carcinoma.
Authors
Affiliations (9)
Affiliations (9)
- Department of Artificial Intelligence, School of Computing, Yonsei University, Seoul, Republic of Korea.
- Yonsei Institute for Digital Health, Yonsei University, Seoul, Republic of Korea.
- Department of Radiation Oncology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea.
- Department of Radiation Oncology, Seoul Metropolitan Government-Seoul National University Boramae Medical Center, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Cancer Research Institute, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Department of Pathology, Seoul National University Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea.
- Department of Radiation Oncology, Seoul National University Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea.
- H-Data Strategy Center, Hallym University Medical Center, Chuncheon, Republic of Korea.
- Department of Biomedical Informatics, Hallym University Chuncheon Sacred Heart Hospital, Chuncheon, Republic of Korea.
Abstract
We aimed to develop models for predicting programmed death-ligand 1 (PD-L1) expression following neoadjuvant chemoradiotherapy (nCRT) in esophageal squamous cell carcinoma (ESCC) using a multifaceted approach that integrates traditional radiomics, deep learning, and machine learning. Pre- and post-nCRT computed tomography (CT) images obtained before surgery and clinical data were retrieved for 101 patients with locally advanced ESCC who underwent nCRT followed by radical surgery. Post-nCRT PD-L1 expression levels, which were obtained from the surgical specimens, were categorized into three groups: < 1%, 1%-10%, and > 10%. Tumor regions on both pre- and post-nCRT CT scans were segmented for radiomic analysis. Radiomics features were extracted and selected prior to modeling using multinomial logistic regression. Predictive models were subsequently developed using machine learning algorithms that integrated radiomics features, clinical variables, and outputs from deep learning architectures, such as Residual Network (ResNet) and TabNet. Using logistic regression, six radiomics features were selected from an initial pool of 1873. Key clinical factors, including radiation technique and total radiation dose, were also identified as predictors. The combined clinical-radiomics model achieved the highest weighted AUROC of 0.8214. However, its improvement over the clinical-only model was small (AUROC, 0.8214 vs. 0.8171), indicating that radiomics provided limited incremental discrimination beyond clinical variables in this cohort. The combined clinical-radiomics model showed feasible performance for assessing post-nCRT PD-L1 expression in patients with locally advanced ESCC, although the incremental discriminative value of radiomics over clinical variables alone was small. Because this study predicted PD-L1 expression rather than actual response to immune checkpoint inhibitors, these findings should be interpreted as hypothesis-generating and require further external validation before clinical application.