CT-based deep learning signatures associated with transcriptomic heterogeneity and combined with nutritional biomarkers improve prediction of 3-year overall survival in esophageal squamous cell carcinoma.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Beijing Friendship Hospital, Capital Medical University, Beijing, China.
- Department of Radiology, Nanjing Pukou People's Hospital, Liangjiang Hospital, Southeast University, Nanjing, PR China.
- Department of Medical Imaging Center, The Affiliated Huaian NO.1 People's Hospital of Nanjing Medical University, Huaian, PR China.
- Department of Medical Imaging, Huaian Hospital Affiliated to Xuzhou Medical University, Huaian, China.
- Department of Radiology, Beijing Friendship Hospital, Capital Medical University, Beijing, China. [email protected].
- Department of Radiology, Beijing Friendship Hospital, Capital Medical University, Beijing, China. [email protected].
Abstract
Deep learning signatures (DLS) extracted from CT images can noninvasively reflect tumor heterogeneity and have shown promise in prognostic modeling for esophageal squamous cell carcinoma (ESCC). To develop and validate a CT-based DL model combined with nutritional biomarkers to predict 3-year overall survival (OS) in ESCC, and to investigate transcriptomic differences between DLS-based risk groups. This retrospective multicenter study included 662 postoperative ESCC patients from three hospitals and 16 additional patients from The Cancer Genome Atlas (TCGA). DL features extraction from CT images based on the Crossformer architecture. Skeletal muscle index was measured at the L3 vertebra to assess low skeletal muscle mass (LSMM). Cox regression was used to build clinical, DL, and combined models. Model performance was evaluated using the concordance index (C-index). Transcriptomic analysis of the TCGA cohort was performed to identify metabolic pathway differences between DLS-based risk groups. The DL model achieved a C-index of 0.743 (95% CI: 0.683-0.803) in the internal validation cohort and 0.692 (95% CI: 0.576-0.809) in the external cohort. Pathological T and N stages, Neuroaggression, Vascular invasion, and LSMM were identified as independent clinical predictors. The combined model achieved a C-index of 0.753 (95% CI: 0.697-0.808) internally and 0.725 (95% CI: 0.613-0.838) externally. DLS-based risk stratification revealed significant differences in metabolic activity between groups, supporting its biological relevance. The combined model enables preoperative OS prediction in ESCC. DLS-based stratification reflects transcriptomic metabolic heterogeneity and enhances the biological interpretability of imaging features. This study developed a CT-based DLS and combined it with nutritional markers for prognostic modeling in ESCC. Transcriptomic analysis of DLS-based groups revealed metabolic heterogeneity, enhancing the biological interpretability of the DL model. A combined DLS and nutritional model enables individualized preoperative survival prediction in ESCC. DLS-based risk groups defined by the DLS exhibited transcriptomic differences in key metabolic pathways, revealing biological underpinnings of imaging-based phenotypes. Attention map visualization revealed consistent spatial focus on morphologically distinct tumor regions, enhancing the interpretability of deep learning predictions.