Intratumoral and peritumoral radiomics for forecasting microsatellite status in gastric cancer: a multicenter study

• Published in: BMC cancer Vol. 25; no. 1; pp. 66 - 13
• Main Authors: Xiao, Yunzhou, Zhu, Jianping, Xie, Huanhuan, Wang, Zhongchu, Huang, Zhaohai, Su, Miaoguang
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 11.01.2025; BioMed Central; BMC
• Subjects: Adjuvants; Adult; Aged; Antigens; Biopsy; Cancer therapies; Complications and side effects; Computed tomography; Correlation analysis; Diagnostic imaging; Female; Gastric cancer; Genetic aspects; Health aspects; Hemoglobin; Hospitals; Humans; Machine Learning; Male; Medical imaging; Medical prognosis; Metastasis; Microsatellite Instability; Middle Aged; Patients; Peritumoral; Prognosis; Radiomics; Regression analysis; Retrospective Studies; ROC Curve; Scanners; Stomach cancer; Stomach Neoplasms - diagnostic imaging; Stomach Neoplasms - genetics; Stomach Neoplasms - pathology; Support Vector Machine; Surgery; Tomography, X-Ray Computed - methods; Tumors; China; Peritumoral; Microsatellite instability; Gastric cancer; Machine learning; Radiomics
• ISSN: 1471-2407; 1471-2407
• DOI: 10.1186/s12885-025-13450-3

This investigation attempted to examine the effectiveness of CT-derived peritumoral and intratumoral radiomics in forecasting microsatellite instability (MSI) status preoperatively among gastric cancer (GC) patients. A retrospective analysis was performed on GC patients from February 2019 to December 2023 across three healthcare institutions. 364 patients (including 41 microsatellite instability-high (MSI-H) and 323 microsatellite instability-low/stable (MSI-L/S)) were stratified into a training set (n = 202), an internal validation set (n = 84), and an external validation set (n = 78). Radiomics features were obtained from both the intratumoral region (IR) and the intratumoral plus 3-mm peritumoral region (IPR) on preoperative contrast-enhanced CT images. After standardizing and reducing the dimensionality of these features, six radiomic models were constructed utilizing three machine learning techniques: Support Vector Machine (SVM), Linear Support Vector Classification (LinearSVC), and Logistic Regression (LR). The optimal model was determined by evaluating the Receiver Operating Characteristic (ROC) curve's Area Under the Curve (AUC), and the radiomics score (Radscore) was computed. A clinical model was developed using clinical characteristics and CT semantic features, with the Radscore integrated to create a combined model. Used ROC curves, calibration plots, and Decision Curve Analysis (DCA) to assess the performance of radiomics, clinical, and combined models. The LinearSVC model using the IPR achieved the highest AUC of 0.802 in the external validation set. The combined model yielded superior AUCs in internal and external validation sets (0.891 and 0.856) in comparison to clinical model [(0.724, P = 0.193) and (0.655, P = 0.072)] and radiomics model [(0.826, P = 0.160) and (0.802, P = 0.068)]. Furthermore, results from calibration and DCA underscored the model's suitability and clinical relevance. The combined model, which integrates IPR radiomics with clinical characteristics, accurately predicts MSI status and supports the development of personalized treatment strategies.