A subscription to JoVE is required to view this content. Sign in or start your free trial.
Research Article
A CT-based radiomics nomogram was developed to accurately differentiate gastric stromal tumors from gastric leiomyomas, enabling noninvasive and individualized preoperative diagnosis with high predictive performance.
This protocol describes the development and validation of a computed tomography (CT)-based radiomics nomogram for the noninvasive preoperative differentiation of gastric stromal tumors (GISTs) and gastric leiomyomas (GLMs), two gastric submucosal lesions with distinct therapeutic strategies and prognostic implications. A retrospective cohort of 172 patients with pathologically confirmed GISTs or GLMs who underwent contrast-enhanced CT within 30 days before surgery was analyzed. Patients were randomly assigned to a training cohort (n = 120) and a validation cohort (n = 52). Demographic variables, CT morphological characteristics, and quantitative radiomic features extracted from manually delineated regions of interest were systematically evaluated. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression with cross-validation. The final predictive model incorporated age, tumor location, enhancement pattern, and the radiomic feature NGLDM_Busyness, which reflects intratumoral texture heterogeneity. These variables were integrated to construct an individualized nomogram for clinical use. Model performance was assessed using receiver operating characteristic analysis, calibration curves, and decision curve analysis. The nomogram demonstrated strong discriminative ability, good calibration, and favorable clinical utility in both the training and validation cohorts, supporting its potential value as a noninvasive tool for individualized preoperative diagnosis.
Gastric stromal tumors (GISTs), which are the most common gastric submucosal tumors (SMTs), tend to be considered potentially malignant regardless of their size, with those exhibiting a malignant clinical course accounting for approximately 10% to 30% of cases1,2. According to the NIH consensus classification3, the clinical malignancy risk of GISTs increases progressively across the strata: very low, low, intermediate, and high risk, which are defined by tumor size and mitotic count. Early diagnosis and early surgical resection are recommended for patients with GISTs4. In contrast, as a benign neoplasm with a rare probability of metastasizing and apparently the most frequent myogenic tumors of gastric SMTs, gastric leiomyomas (GLMs) require conservative observation or minimally invasive treatments instead of surgery, except for cases in which the lesions are larger than 5 cm5,6. Thus, it is of great importance to distinguish GISTs from GLMs smaller than 5 cm because of their substantially different therapeutic and prognostic implications. This distinction is particularly challenging for small lesions, as they often share overlapping imaging features and lack definitive clinical symptoms7,8.
Endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA), as the most mature technique relating to pretherapeutic identification in gastric SMTs, has been proven to provide a safe and accurate pathological diagnosis of GISTs7. Unfortunately, the ability of EUS-FNA to extract tissue samples is severely limited, which appears to reduce the accuracy of its diagnosis9. Moreover, the technique is considered to be invasive. Computed tomography is considered an optimal and economical imaging tool for the preoperative diagnosis of gastrointestinal tract tumors10. However, differentiating between GISTs and GLMs preoperatively remains challenging, given the complexity of their similar clinical symptoms and CT appearances10.
Radiomics represents a promising tool with the ability to extract quantitative imaging features from radiographic images automatically, allowing the objective quantification of the heterogeneity of tumors11. In previous studies, radiomics models have been developed primarily for risk stratification of GISTs based on CT or MRI, demonstrating their potential in evaluating malignant potential12,13,14. However, studies focused specifically on the differential diagnosis between GISTs and GLMs remain relatively rare. To present knowledge, no prior CT-based radiomics nomogram has been developed for the preoperative differentiation of GISTs from GLMs. Therefore, this study constructs a preoperative prediction nomogram using morphological features of CT images and radiomics parameters to address the abovementioned challenges.
This research was approved by the ethics committee of the hospital (protocol number: 1612167-18). Given the retrospective nature of the study, the requirement for informed consent was waived. Patients with gastric stromal tumors (GISTs) or gastric leiomyomas (GLMs) were identified from institutional records between January 2017 and July 2022, all with postoperative pathological confirmation. The reagents and the equipment used are listed in the Table of Materials.
1. Patient details
Eligible patients are enrolled if they meet the following criteria: (1) gastric lesions with a maximum diameter ≤5 cm; (2) availability of complete pathological diagnosis; (3) undergoing radical surgery with curative intent; and (4) having undergone contrast-enhanced CT within 30 days prior to surgery. Patients with a history of malignant tumors or preoperative therapy were excluded. Therefore, 110 GIST patients and 62 with GLMs were enrolled in the present study. A population flowchart was demonstrated (Figure 1). A total of 172 patients (110 GISTs, 62 GLMs; 82 males, 90 females) were included, with a mean age of 55.05 ± 12.42 years (range: 22-79). Patients were randomly assigned to a training set (n = 120) and a validation set (n = 52) using simple randomization without stratification. Baseline demographic data are summarized in Table 1.
2. CT image analysis
CT scanning within the present research was performed on 64-slice CT scanners. Every patient was required to fast for at least 6 h. Approximately 30 min before the CT scan, each patient was instructed to orally ingest 500-800 mL of tap water, and 1000 mL of water was requested to be taken immediately before the scan to distend the stomach. All patients in the study were positioned in the supine position to prevent artefacts caused by air in the stomach, while ensuring that all lesion areas were covered. Patients were informed of exposure to ionizing radiation as part of the clinical imaging procedure. Prior to contrast administration, contraindications to iodinated contrast medium, including known allergy and renal insufficiency (eGFR <30 mL/min/1.73 m²), were screened. Patients were monitored for at least 15 min post-injection for acute adverse reactions. A standard reconstruction algorithm was apparently utilized, and the scanning process strictly followed these specifications: 120 kV tube voltage; 250-300 mA tube current; 1.5 mm slice thickness and interval. A soft-tissue reconstruction kernel (B30f) was used. Subsequently, during the enhanced scanning process, 80-120 mL iodinated contrast agent was administered, with an injection rate of 3.0 mL/s intravenously. Arterial, portal, and delayed phases should be collected during approximately 30 s, 60 s, and 120 s after the injection time. For radiomic analysis, images were preprocessed by resampling to an isotropic voxel size of 1.5 × 1.5 × 1.5 mm³ and intensity discretization using a fixed bin width of 25 Hounsfield units (HU).
3. Radiomic feature extraction
As for radiomic features, LIFEx software (version 4.90; www.lifexsoft.org) was used to extract features from the region of interest (ROI). All regions of interest were delineated on portal venous phase CT images only, which were selected for radiomic analysis to ensure consistency and feature stability. For each patient, the ROI was delineated on all slices containing the lesion in the transverse plane, with sagittal and coronal views used to ensure accurate volume coverage (Figure 2). ROIs were independently drawn by two radiologists without consensus during initial segmentation to enable reproducibility assessment. The ROI represented a two-dimensional (2D) single-slice contour, not a three-dimensional volume. All tumor areas were included when delineating ROIs, and calcifications were excluded based on density thresholds in the LIFEx software setting. Visual quality control was performed to ensure adequate gastric distension and complete inclusion of the lesion within the ROI. Intra- and interclass correlation coefficients (ICCs) were calculated to assess reproducibility of radiomic feature extraction; features with ICC > 0.75 were retained. Through the automatic algorithm of LIFEx software, 37 radiomic features were obtained: 5 histogram parameters, 7 grey-level co-occurrence matrix (GLCM) parameters, 11 grey-level run length matrix (GLRLM) parameters, 3 neighbourhood grey-level difference matrix (NGLDM) parameters, and 11 grey-level zone length matrix (GLZLM) parameters. Feature extraction was performed via the menu path: Segmentation → Radiomics → Compute. Image preprocessing included resampling to an isotropic voxel size of 1.5 × 1.5 × 1.5 mm3 and intensity discretization using a fixed bin width of 25 Hounsfield units (HU). No feature scaling or normalization was applied prior to modeling.
4. Statistics
Statistical analyses were performed using R software (version 3.6.0; https://www.r-project.org). Continuous variables were compared between groups using independent t-tests; categorical variables were assessed with χ² tests. A two-sided P value < 0.05 was considered statistically significant.
5. Establishment of prediction nomogram
Univariate analysis was conducted using Pearson's correlation to screen demographic, CT morphological, and radiomic characteristics for association with diagnosis (GIST vs. GLM). Multivariate modeling employed least absolute shrinkage and selection operator (LASSO) regression implemented via the glmnet package (cv.glmnet function) with 10-fold cross-validation15. The optimal regularization parameter λ was selected at lambda.1se to favor a more parsimonious model. The final predictors, age, tumor location, enhancement pattern, and NGLDM_Busyness, were linearly combined using their respective coefficients to generate a prediction score (Prescore). Among all extracted radiomic features, only NGLDM_Busyness retained a non-zero coefficient after LASSO penalization, suggesting that other radiomic features were either redundant or provided limited additional discriminative value. A nomogram was constructed based on the multivariate model using the nomogram function from the rms package.
6. Prediction effectiveness and validation of the nomogram
Predictive performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy in both training and validation cohorts. Calibration was assessed via bootstrapping resampling (1000 iterations) and Hosmer-Lemeshow goodness-of-fit test; calibration curves were plotted accordingly. A P value < 0.05 indicated a significant deviation from ideal calibration. The concordance index (C-index) was used to quantify discriminative reliability. Clinical utility was evaluated by decision curve analysis (DCA).
Demographic and clinical features
A total of 172 patients with gastric submucosal tumors (110 GISTs, 62 GLMs) were included. Demographic and clinical features are summarized in Table 1, Table 2 and Table 3. This analysis aimed to identify baseline differences between GISTs and GLMs to inform subsequent model development. GISTs were more likely to occur in the gastric body in elderly patients with a moderate enhancement pattern, whereas GLMs were more frequent at the cardia in younger patients with slight enhancement (p < 0.05). The Hounsfield unit (HU) on plain CT scan was significantly lower in GIST patients than in GLM patients (p < 0.05). No significant differences were observed in sex, symptoms, tumor history, family history, size, growth pattern, shape, CT value on enhanced scanning, lesion boundary, or presence of cystic changes or calcifications (p > 0.05). Continuous variables were compared using independent t-tests after confirming normality via the Shapiro-Wilk test; categorical variables were assessed with χ² tests.
Radiomic feature reproducibility and selection
Ten CT morphological features and 37 radiomic characteristics were extracted. To evaluate inter-observer reproducibility, a prerequisite for reliable radiomic analysis, two radiologists independently delineated ROIs, and intra- and interclass correlation coefficients (ICCs) were calculated. Inter-observer reproducibility was assessed using intra- and interclass correlation coefficients (ICCs). All extracted radiomic features had ICC values greater than 0.85, indicating excellent agreement between the two radiologists. The mean ICC was 0.91, with a range of 0.86 to 0.94. Given this reproducibility, the mean values of the two readers' measurements were used in subsequent analyses. Cross-correlation matrices revealed substantial interrelationships among the 47 candidate features (Figure 3).
Establishment of the prediction model and nomogram
The goal of this step was to develop a parsimonious model for differentiating GISTs from GLMs using clinically and radiomically relevant predictors. LASSO regression selected four key features: age, tumor location, enhancement pattern, and the radiomic parameter NGLDM_Busyness (Figure 4). The prediction score (Prescore) was calculated as:
Prescore = 304.924 − 3.985 × age − 52.115 × location − 11.325 × enhancement pattern − 135.104 × NGLDM_Busyness.
Patients with GLMs had significantly higher Prescores than those with GISTs (p < 0.05) (Figure 5). A nomogram was constructed based on these four predictors (Figure 6). The model achieved high discriminative performance, with AUCs of 0.92 (95% CI: 0.87-0.96) in the training set and 0.89 (95% CI: 0.81-0.95) in the validation set (Table 4).
Validation and clinical utility of the nomogram
This section evaluates the calibration and clinical applicability of the final nomogram in both the training and validation cohorts. Calibration analysis revealed close agreement between the nomogram-predicted probabilities and the actual observed outcomes in both groups. Decision curve analysis further demonstrated that the nomogram provides a positive net clinical benefit across a wide range of threshold probabilities (greater than 10%), supporting its potential utility in real-world clinical decision-making.
Summary of key findings
The study identified age, tumor location, enhancement pattern, and NGLDM_Busyness as key predictors for differentiating gastric stromal tumors from leiomyomas. The resulting nomogram demonstrated strong discrimination, good calibration, and favorable clinical utility in both training and validation sets.
DATA AVAILABILITY:
All raw data used and analyzed in this study are provided in Supplementary File 1 and Supplementary File 2.
Figure 1: A population flowchart. GIST, gastric stromal tumor; GLM, gastric leiomyoma. Please click here to view a larger version of this figure.
Figure 2: Portal venous phase CT images of a 35-year-old man (A-C) and a 41-year-old male patient (D-F). (A,D) Axial sections showing the region of interest (ROI) manually delineated on the slice with the largest tumor cross-sectional area. (B,E) Coronal reformatted views displaying the same ROI as visualized in the anteroposterior dimension. (C,F) Sagittal reformatted views displaying the same ROI as visualized in the left-right dimension. Green contours represent the three-dimensional tumor volume used for radiomic feature extraction, reconstructed from multiplanar segmentation. Please click here to view a larger version of this figure.
Figure 3: Cross-correlation matrices. The depth of color indicates the intensity of the correlation between covariates. The darker the color, the higher the correlation is. Blue represents positive correlation, and red represents negative correlation. Please click here to view a larger version of this figure.
Figure 4: Key characteristics selection for predictive model. The vertical axis indicates the model misclassification rate, and the horizontal axis shows log (λ). The two vertical dashed lines represent one standard deviation on each side of the minimum value, corresponding to the chosen variables that better fit the models. Please click here to view a larger version of this figure.
Figure 5: Pre-scores of the model for each patient. Please click here to view a larger version of this figure.
Figure 6: ROC curves of the prediction nomogram in the training (A) and validation (B) cohorts. AUC values were 0.969 (95% CI: 0.905-0.974) and 0.943 (95% CI: 0.905-0.949), respectively. Optimal cutoffs were 0.417 and 0.297. Please click here to view a larger version of this figure.
Table 1: Baseline clinical characteristics of patients with gastrointestinal stromal tumors (GISTs) and gastric leiomyomas (GLMs) in the training and validation cohorts. Please click here to download this Table.
Table 2: Comparison of CT imaging features between patients with gastrointestinal stromal tumors (GISTs) and gastric leiomyomas (GLMs) in the training and validation cohorts. Please click here to download this Table.
Table 3: Comparison of CT imaging features, including calcification, cystoid changes, Hounsfield units (HU), and enhancement patterns between patients with gastrointestinal stromal tumors (GISTs) and gastric leiomyomas (GLMs) in the training and validation cohorts. Please click here to download this Table.
Table 4: Diagnostic performance of the radiomics-based prediction model in distinguishing gastrointestinal stromal tumors (GISTs) from gastric leiomyomas (GLMs) in the training and validation cohorts. Please click here to download this Table.
Supplementary File 1: Raw data for the training dataset. Please click here to download this File.
Supplementary File 2: Raw data for validation. Please click here to download this File.
Discriminating between gastrointestinal stromal tumors (GISTs) and gastric leiomyomas (GLMs) is clinically important because it directly affects treatment decisions and patient outcomes. A nomogram was developed and validated as a predictive nomogram that incorporates one demographic feature (age), two CT signs (location and enhancement pattern), and one radiomic parameter (NGLDM_Busyness). This model can be readily applied preoperatively in clinical practice. This study addresses a current gap in noninvasive preoperative differentiation of gastric submucosal tumors by providing a practical alternative to methods with limited diagnostic yield or invasiveness.
In the present analysis, the mean age of patients with GISTs was 59.69 ± 8.90 years in the training cohort and 60.94 ± 9.27 years in the validation cohort, which aligns with findings from He et al.16 and Wang et al.17. GIST patients were older than GLM patients, consistent with previous reports18. Levy et al.19and Miettinen et al.20 reported that GLMs are exclusively located in the cardia, whereas GISTs most commonly arise in the gastric body, a pattern confirmed in the present data. The results also showed that GISTs showed greater contrast enhancement than GLMs, a result consistent with Wang et al.21. This difference likely reflects the higher microvessel density in GISTs, which are tumors with malignant potential. Although their vascular supply is lower than that of overt carcinomas, it exceeds that of benign leiomyomas, explaining the moderate enhancement observed in the present cohort.
Among radiomic features, NGLDM_Busyness, a second-order texture metric derived from the neighborhood grey-level difference matrix (NGLDM), quantifies local grey-level variations and correlates with three-dimensional tumor heterogeneity22. GIST patients exhibited significantly higher NGLDM_Busyness values than GLM patients, indicating greater intratumoral heterogeneity. This finding aligns with prior studies reporting more frequent necrosis and calcification in GISTs than in GLMs21,23, features known to contribute to textural complexity and malignancy-associated heterogeneity24. However, this study did not observe significant differences in calcification or cystic changes between the two groups, likely because this study restricted this analysis to GISTs smaller than 5 cm. Larger prospective studies should further investigate this phenomenon.
Endoscopic ultrasound-guided fine-needle aspiration biopsy (EUS-FNAB) is the clinical gold standard for diagnosing subepithelial lesions (SELs), but its diagnostic yield ranges only from 60% to 85%25 and carries procedural risks. Recent advances in artificial intelligence have enabled AI-augmented EUS, which achieves a sensitivity of 92%, specificity of 80%, accuracy of 46%, and an AUC of 0.92 (95% CI: 0.90-0.94) [26]. In comparison, present CT-based nomogram demonstrated superior performance: in the training set, AUC = 0.969 (95% CI: 0.926-1.000), sensitivity = 90.58%, specificity = 97.43%, and accuracy = 95.06%; in the validation set, AUC = 0.943 (95% CI: 0.8725-1.000), sensitivity = 90.51%, specificity = 94.90%, and accuracy = 93.34%26. Although EUS-FNA remains the reference standard, a noninvasive CT-based model offers a complementary tool that can be integrated into routine radiology workflows, especially for patients who are poor candidates for endoscopy. Alternative approaches, such as MRI or deep learning-based image analysis, were not evaluated here but warrant investigation in future work.
This study has several limitations. First, its single-center retrospective design introduces potential selection bias; multicenter prospective validation is therefore essential. Furthermore, all imaging was acquired using a single CT scanner type, which may limit generalizability across institutions with different acquisition protocols or reconstruction algorithms. Second, this cohort included only patients who underwent preoperative CT before curative-intent surgery, which may overrepresent malignant or symptomatic cases. Third, this study focused exclusively on GISTs and GLMs, the two most common and clinically relevant gastric submucosal tumors, but a broader model incorporating additional subtypes would enhance clinical applicability. Potential applications of nomograms include integration into PACS for real-time decision support and use in preoperative risk stratification to guide surgical planning.
In conclusion, this study established and validated a convenient, reliable, and predominantly noninvasive nomogram that combines demographic, imaging, and radiomic features to differentiate GISTs from GLMs. This tool can support precise clinical decision-making. Future work should prioritize multicenter external validation, multimodal imaging fusion (e.g., combining CT with EUS features), and expansion to additional submucosal tumor subtypes to maximize clinical utility.
The authors declare that they have no competing interests.
None.
| Contrast-enhanced CT scanner | Siemens Healthineers | SOMATOM Definition AS+ | 64-slice CT scanner; typical in Chinese tertiary hospitals during 2017–2022. Scanning parameters: 120 kV, 250–300 mA, 1.5 mm slice thickness, B30f kernel. |
| glmnet package (R) | CRAN | Version 4.0 (compatible with R 3.6.0) | LASSO regression with 10-fold cross-validation. |
| Iodinated contrast agent | Bayer AG | Ultravist 370 (Iopromide 370 mg I/mL) | Administered IV at 80–120 mL, 3.0 mL/s. Widely used non-ionic contrast agent in China for abdominal CT. |
| LIFEx software | www.lifexsoft.org | Version 4.90 | Used for radiomic feature extraction from CT ROIs. |
| R software | The R Foundation | Version 3.6.0 | For statistical analysis and modeling (https://www.r-project.org). |
| rms package (R) | CRAN | Version 5.1-4 (compatible with R 3.6.0) | Nomogram construction and calibration. |
