Integrated Assessment of Sarcopenia in Patients with Gastric Cancer Using Deep Learning and Radiomics

BACKGROUND: Sarcopenia is a common condition in patients with gastric cancer that is associated with poor survival and adverse clinical outcomes. Conventional diagnostic approaches are time-consuming, limiting large-scale clinical implementation. This study aimed to develop and validate a multimodal model that integrates clinical variables, radiomics features and deep learning features. METHODS: A total of 1067 patients with gastric cancer from two medical centres were included. The diagnosis of sarcopenia required low grip strength and a low skeletal muscle index, whereas severe sarcopenia additionally required low gait speed. Radiomics features were extracted from computed tomography images at the third lumbar vertebral level to construct radiomics models. The 2D and 2.5D deep learning models were developed using the ResNet50 architecture. A transformer-based multimodal model integrating clinical variables, radiomics features and 2.5D deep learning features was developed and termed the sarcopenia model, while an XGBoost fusion model integrating the same three modalities was constructed as a comparator. Model performance was evaluated in the training, validation and external test sets using receiver operating characteristic and decision curve analyses. RESULTS: Among the 1067 patients, 782 (73.3%) were male and 285 (26.7%) were female, with no significant difference in gender distribution across the training, validation and external test sets (p = 0.34). Kaplan-Meier survival analysis demonstrated that the overall survival of patients with sarcopenia was significantly worse than that of patients without sarcopenia (p < 0.05). Among single-modality models, the radiomics model achieved AUCs of 0.85 and 0.80 in the validation and external test sets, respectively. The 2.5D deep learning model outperformed the 2D model, yielding AUCs of 0.81 and 0.83 in the validation and external test sets. The XGBoost model achieved AUCs of 0.85 and 0.86 in the validation and external test sets, respectively. In the sarcopenia model, the AUCs for identifying sarcopenia were 0.95, 0.87 and 0.89 in the training, validation and external test sets, respectively, while those for identifying severe sarcopenia were 0.93, 0.85 and 0.84. CONCLUSIONS: The sarcopenia model integrating clinical variables, radiomics features and deep learning features enables the accurate identification of sarcopenia and severe sarcopenia in patients with gastric cancer and may support clinical screening and risk stratification.