A deep learning-based radiomic nomogram derived from visceral fat for early prediction of gastrointestinal stromal tumor risk grade.

Authors

Abstract

Gastrointestinal stromal tumors (GISTs) show substantial heterogeneity and are classified into distinct risk categories requiring different treatment strategies. Therefore, reliable preoperative risk stratification is crucial for treatment planning. Due to its inherent contrast, visceral adipose tissue can be reliably segmented and its volume of interest (VOI) obtained even on non-contrast CT scans. Therefore, this study aimed to develop a deep-learning-based radiomics nomogram (DLRN) that integrates visceral adipose features extracted from non-contrast computed tomography (CT). The DLRN is designed to provide a convenient tool for the preoperative prediction of GIST risk grades, including for patients for whom contrast-enhanced imaging is not feasible. A total of 211 patients with histologically confirmed GISTs from two institutions were included. The derivation cohort from Institution A (<i>n</i> = 158) was randomly divided into a training cohort (<i>n</i> = 110) and an internal validation cohort (<i>n</i> = 48) at a 7:3 ratio. An independent external test cohort from Institution B (<i>n</i> = 53) was used for external validation. Visceral fat features were extracted from non-contrast CT images, and the DLRN was constructed for preoperative risk grading. Model performance was compared with clinical, traditional radiomics, deep learning radiomics, and feature-fusion models. Among the evaluated models, the DLRN showed favorable discrimination in the derivation cohort and exploratory performance in the external test cohort. In the derivation cohort, it achieved an area under the curve (AUC) of 0.936 [95% confidence interval (CI): 0.8907-0.9812], accuracy of 0.873, sensitivity of 0.811, and specificity of 0.904. In the external test cohort, the AUC was 0.862 (95% CI: 0.6216-1.0000), with an accuracy of 0.925, a sensitivity of 0.833, and a specificity of 0.936. Decision curve analysis demonstrated that the DLRN provided a higher net clinical benefit than all comparison models in both datasets. Calibration curves showed the agreement between predicted probabilities and observed outcomes, while DeLong tests were used for pairwise AUC comparisons between models. The proposed DLRN, integrating visceral fat-derived radiomics and clinical variables, may provide a non-invasive adjunct for preoperative GIST risk stratification on non-contrast CT. Further validation in larger multicentre cohorts is still needed.

Tags

Topics