Patients
This retrospective study was approved by the Institutional Review Board of by the Institutional Review Board of West China Hospital affiliated with Sichuan University (No. 2020–249), and the qualification for informed consent was waived. All procedures performed in this study involving human participants conformed with the ethical guidelines of the Declaration of Helsinki (as revised in 2013). Data from patients who met the following criteria at West China Hospital of Sichuan University from May 2010 to December 2022 were retrospectively analyzed. Inclusion criteria: i, those with pathology confirmed as siGISTs postoperatively at West China Hospital, and tissue specimens were subjected to genetic testing; ii, those with complete CECT imaging data within one month preoperatively; iii, those with a time interval of < 4 weeks between CECT and surgery outcomes. Exclusion criteria: i, history of GISTs, or other primary malignant tumors; ii, preceding records of any treatment, such as surgery, chemotherapy, and radiotherapy, before CECT examination; iii, insufficient image data to meet the needs of diagnosis and tumor segmentation (no enhanced image, incomplete image scan range, severe artifact, etc.); iv, tumor rupture, target lesion not found on CECT, or inability to accurately segment the tumor; v, genetic tests confirmed as non-KIT gene mutations siGISTs. A total of 152 patients with siGISTs were initially eligible. Among these, 9 patients were excluded because of a history of other tumors; 33 because of a preceding history of surgery or treatment; 12 because of insufficient or inappropriate CECT images; and 7 because of tumor rupture, invisible, or too small target lesions on CECT (Exclusive total n = 61). Finally, this study enrolled 91 with patients siGISTs (Fig. 1).
CECT scanning parameters
Patients ate a light diet the day before the examination and fasted for 8 h on the day of the examination. The patient ingested 600–1000 ml of warm wate by mouth 40 min to 1 h before the scan. The following multiple CT scanners were performed: Philips Brilliance 64; Siemens SOMATOM Definition AS+; Siemens SOMATOM Definition Flash. The scan parameters were as follows: tube voltage of 120 kV, tube current of 120–160 mAs (plain phase) or 140–215 mAs (enhanced phases), layer thickness of 1–5 mm, layer spacing of 5 mm, field of view of 30–50 cm, matrix of 512 × 512, and pitch of 0.9–1.2. The patient was scanned from the top of the diaphragm to the superior border of the pubic symphysis in the supine position with the arms extended on the side of the head. A conventional plain scan was first performed, followed by an enhanced scan. A high-pressure double-ended syringe was used to inject contrast agent (1.0 mL/kg, 3.0–4.5 mL/s) and then flushed with 20–30 mL of saline through the elbow vein. The arterial and venous phase images were acquired at 50 s and 80 s after contrast injection, respectively.
CECT Imaging Feature Analysis
All CECT images of the included patients were exported in Digital Imaging and Communications in Medicine format in the Picture Archiving and Communication System and presented to Syngo Imaging Workplaces (VersionVB35A, Siemens AG, Erlangen, Germany). Two experienced abdominal radiologists (Xi-jiao Liu and Cai-wei Yang, with 11, and 6 years of abdominal imaging experience, respectively) who were blinded to clinical records, pathological results, and each other’s conclusions, performed and recorded image analysis and the consensus was reached after discussion in case of any disagreement. Observations: (1) tumor size; (2) tumor margin, which is defined as well-defined or ill-defined; (3) tumor contour, including regular or irregular shape; (4) growth pattern, including endophytic (the tumor attached to the inner edge of the gastrointestinal wall and completely confined to the gastrointestinal lumen without growth outside the gastrointestinal lumen), exophytic (the tumor growing outside the gastrointestinal wall without protruding into the lumen), or mixed growth type (both the above conditions); (5) tumor attenuation, which is classified as hypodensity, slightly hypodensity, iso-density, and hyper-density, as a comparison of the same level paraspinal muscle on plain scan. (6) enhancement pattern, which is categorized as homogenous and heterogenous on enhanced scans; (7) enhancement level, which is grouped as mild, moderate, and marked. The level was based on tumor CT attenuation difference value between the enhanced arterial and plain phases (differences of ≤ 20 Hounsfield unit [HU] were defined as mild; 20–40 HU were defined as moderate, and ≥ 40 HU were defined as marked); (8) the presence of necrosis, described as a region with CT attenuation value of ≤ 20 HU in each phase and with its enhancement level of ≤ 10 HU; (9) calcification; (10) superficial ulceration, which is named as a discontinuity or break in the intestinal mucous membrane; (11) gas density in tumor; (12) cystic change, which referred to a water-like area (CT attenuation value ≤ 10 HU) with ring-shaped enhancement with a thin or irregularly thick wall enhancement; The main difference between necrosis and cystic change is that the signs of cystic change are based on areas of watery CT density and must be accompanied by cystic wall tissues. (13) enlarged vessels, which is depicted as veins or arteries around the tumor are markedly enlarged or filled; (14) fatty infiltration in peripheral mesangial, referred to increased fat density and blurred edges in the adjacent mesentery of the tumor, is a sign that it may be more aggressive; (15) adjacent organ invasion; (16) metastasis; (17) lymphadenopathy.
CECT Radiomics Analysis and Model Construction
Image segmentation: only venous phase images were selected for radiogenomics analysis in this study to ensure the accuracy and consistency of the results, considering the tumor sites of siGISTs, and all tumors were most clearly displayed in the venous phase. The CECT images (venous phase) were imported into ITK-SNAP (www.itk-snap.org) software, and the regions of interest (ROIs) were delineated independently by the above two readers. Images were preprocessed before delineation, and all venous parenchymal phase images were processed to a slice thickness of 1 mm. Specific principles of ROI delineation were as follows: (1) the distance between the edge of the ROIs of the tumor and the inner side of the edge of the tumor should be 1 mm; (2) delineating the largest axial layer plus a nearby axial layer (peritumor region) to obtain double ROIs of the tumor; (3) included all components of the targeted lesion, including necrosis, calcification, inner gas, and blood vessels; and (4) avoided gas outside the tumor, intestinal wall, and surrounding mesangial adipose tissue.
Feature extraction: The original image of the venous phase and the corresponding delineated ROIs file were implied in AK software (GE Artificial Intelligence KIT) to extract the radiomics parameters, including Original, Ipris, Wavelets, LBP, PLBP, WILBP, CoLIAGe, Shearlets, Gabors, etc. The software automatically extracted the radiomics parameters while completing the discretization step. The radiomics features extracted from the ROIs of 50 tumors were randomly selected for consistency parameter evaluation for inter-individual comparisons. The radiomics features with consistency parameter evaluation values of ≥ 0.75 in intra- and inter-classes correlation coefficients (ICC) simultaneously were included in the future selection and elaboration of the radiomics model.
Data processing, model building, and evaluation of each model: 70% was selected as the training set and 30% as the validation set, and the allocation results were obtained following the random stratified sampling principle. Repeated stratified partitioning was used to reduce the biased selection of a single validation data set. Features with variance < 1.0 were excluded based on the features retained by in the ICC the analysis. Then, the outlier values greater than the third quartile + 2 × quartile distance was converted to the 95th percentile, while values less than the first quartile – 2 × quartile distance was converted to the 10th percentile. The synthetic minority oversampling technique (SMOTE) algorithm, which was used to help overcome the problem of the small sample size with a mild imbalance in the training dataset, was used to select the most predictive radiomics and CECT features from the primary data set(13). Next, all the features were normalized and standardized by the Z-Score method. The importance features were evaluated by random forest (RF) following the mean decrease in Gini calculated for all decision trees in RF, and the top 3 important features were finally retained and used for constructing the RF model (mtry = 1, ntree = 310)(14). The diagnostic performances of the CECT signs model, the radiomics model, and the combined model were evaluated by areas under the curve (AUCs) value in training and validation sets. The area under the receiver operating characteristic (ROC) curve was used for representing comprehensive performance. The criteria are as follows: an AUC value of 0.5–0.7 indicates low diagnostic efficiency, 0.7–0.9 indicate good diagnostic efficiency, and ≥ 0.9 indicate excellent diagnostic efficiency. The specificity, sensitivity, positive predictive values, and negative predictive values were used to represent model performance at a specific model threshold, which was determined by maximizing the Youden index. The Delong test was used to compare the AUC of paired models. Internal validation was estimated with a regular bootstrapping of 1000-times bootstrap samples(15). The Hosmer-Lemeshow test was applied to assess the goodness-of-fit of the model, with P-values of > 0.05 representing the agreement between the observed and predicted values. The model calibration was visualized by the calibration curve analysis, and the clinical net benefit of the model was evaluated by decision curve analysis (DCA).
Statistical Analysis
Tests for normality and homogeneity of variance were performed for quantitative parameters, and differences between groups were compared using independent samples t-test or Mann-Whitney-Wilcoxon test. Qualitative parameters were analyzed using the chi-square test or Fisher’s exact test. Statistical tests were performed using statistical software (Version 19, Chicago, IL, USA) and R software (Version 3.6.3; http://www.Rproject.org). All statistical significance levels were two-sided, and a P-value of ≤ 0.05 was considered statistically significant.