2.1 Patients
The ethics review boards of our hospitals granted their approval for this retrospective research, and the patient's informed consent was not needed.
Between April 2018 and July 2022, data were collected from patients pathologically confirmed with AGC via biopsy or surgery in our hospital. The specific inclusion criteria were as follows: (1) AGC established histologically; (2) visible lesion on DCE-MRI; (3) no anticancer therapy prior to DCE-MRI; and (4) DCE-MRI within two weeks of biopsy or surgery. These exclusion criteria were as follows: (1) significant imaging abnormalities that hampered lesion characterization; (2) maximum tumor diameter of 1 cm; and (3) absence of preoperative clinical data. Ultimately, 103 people were enrolled in the research study (Fig. 1).
2.2 MRI Scanning
Before the MRI, all patients received the following preparation: (1) fasted for 8 hours to allow the gastrointestinal tract to empty. (2) To suppress gastrointestinal motility, 10 mg anisodamine (Hangzhou Minsheng Pharmaceutical Co., LTD., China) was administered intramuscularly 10 minutes before the examination if there were no contraindications (e.g., glaucoma, asthma, or serious heart disease). (3) Patients were given 800–1000 mL of warm water orally 5 minutes before the exam to expand the stomach cavity.
For the MRI studies, a typical 12-channel phased-array body coil was employed in conjunction with a 3.0T MRI scanner (Verio, Siemens, Germany). The patient was lying supine during the examination, and the entire stomach was covered by the scanning field. Following a standard plain scan (T1-weighted image, T2-weighted image), a DCE-MRI scan was needed for all patients. Free breathing is employed during DCE-MRI scans, which are performed utilizing a three-dimensional, radial volumetric interpolated, breath-hold assessment approach. Initially, the following parameters were utilized for multiangle cross-sectional T1WI in the axial plane scan: repeat time: 3.25 ms; echo time: 1.17 ms; FOV: 350 × 284 mm; matrix: 288 × 164; layer thickness: 5 mm; scan at various flip angles (5°, 10°, and 15°) for 6.5 s each, for a total of 19.5 s. The next step employed multiphase dynamic enhanced scanning with the following parameters: the Flip angle was set to 10°, 35 phases were scanned, and the total scanning time was 227.5 s. All other parameters were left at their previous values. In phase 3, a gadolinium contrast material (Omniscan, GE Healthcare, China) was injected through the median elbow vein using a high-pressure injector. The injection dose and rate were set at 0.1 mmol/kg and 3.5 ml/s, respectively. To flush the region, 20 ml of saline was administered at the same flow rate.
2.3 Image Data Analysis and Processing
We used Omni Kinetics (GE Healthcare, China) software to postprocess the DCE-MRI image data of all qualified AGC patients. To rectify ROI displacement brought on by involuntary motions such as patient breathing, three-dimensional nonrigid registration preprocessing was applied to dynamic enhanced sequence images. Next, corrected dynamic enhanced sequence scans and 5°, 10°, and 15° degree multiflip angles are postprocessed using Omni Kinetics software. The arterial input function (AIF) was drawn after choosing the abdominal aorta manually. Three DCE-MRI pseudocolor pictures were manually created by setting the region of interest (ROI) using a pharmacokinetic Tofts model. These measurements included the transfer constant (Ktrans), rate constant (Kep), and extravascular extracellular volume (Ve). Five different types of features were automatically extracted by the software from three perfusion maps of 231 radiomics features, including first-order, histogram, gray-level co-occurrence matrix, Haralick, and run length matrix. Two seasoned radiologists chose and manually defined ROIs on T1WI-enhanced images for the model in a double-blind fashion. Layer by layer, lesions were identified and combined into three-dimensional ROIs for analysis. To prevent cystic alterations, necrosis, and healthy stomach tissue, certain areas were marked off (Fig. 2). The mean value was calculated after three iterations of all calculations.
2.4 Immunohistochemical Staining and Analysis
The expression of CD3+, CD4+, and CD8+ T cells in gastric cancer tissues was examined using immunohistochemistry (IHC). Pathological samples for gastric cancer were obtained through gastroscopic biopsy or surgery. All GC tissues that had been formalin fixed and paraffin coated were sliced into 4-µm-thick slices. Immunohistochemical staining was carried out using mouse anti-CD8 monoclonal antibody (1:200, GT211202, Gene Tech, Shanghai, China), rabbit anti-CD4 monoclonal antibody (1:200, GT219102, Gene Tech, Shanghai, China), or rabbit anti-CD3 monoclonal antibody (1:200, GT219001, Gene Tech, Shanghai, China). Overnight, the portions were kept in a 4°C refrigerator. The samples were then stained with a secondary antibody (K5009, Dako, Beijing, China) and incubated at 37°C for 10 min. Hematoxylin was employed as a counterstain, and diaminobenzidine (DAB) was utilized to designate the antibody. Before being examined under a microscope, sections were made transparent, dried, and mounted. Two knowledgeable pathologists conducted a double-blind examination of the immunohistochemical results. A low-power microscope was used to examine the complete tissue field before five randomly chosen fields were examined using a high-power (X40) microscope (Fig. 3). The tumor tissue and stroma surrounding it, as well as cancer cell nests, were all included in the counting field. Patients were divided into two groups based on the median after CD3+, CD4+, and CD8+ T-cell expression was evaluated based on the average number of positively stained cells, according to an earlier study(Lee et al. 2008).
2.5 Feature Selection
The average value of each extracted radiomics feature was subtracted, its standard deviation was divided by it (a process known as Z score normalization), and all of the original feature values were then transformed into feature values with a 0–1 normal distribution. All features that are extracted may not apply to a particular activity. Therefore, a critical step for achieving the most effective result is to screen out particular features that are most pertinent to this study. In this work, three strategies for dimension reduction were used to eliminate redundant features: the variance threshold, the single variable selection method, and the least absolute shrinkage and selection operator (LASSO) method. Features with less than 0.8 variance are first eliminated by the variance cutoff. A p value is used to assess the link between features and classification outcomes in the SelectKBest method. The screening of all characteristics with a p value less than 0.05 is possible using this univariate feature selection technique. L1 regularization is used in LASSO regression as the cost function, with a maximum of 1000 iterations, to eliminate weakly correlated features and ultimately produce the best feature selection.
2.6 Construction and Validation of Radiomics Models
Because only 103 patients were enrolled, it was impossible to evaluate the robustness of our model using the conventional method of splitting the sample into training and validation groups. Using 10-fold cross-validation, our study evaluated the resilience of the prediction model. The training data were subjected to a 10-fold internal cross-validation. The training data were divided into ten subsets; one subset was used for validation, while the other nine subsets were used for training. The next 10 iterations followed. Models for logistic regression classification were trained using these data. The confusion matrix was used to calculate the model's accuracy, sensitivity, and specificity. Plots of the receiver operating characteristic curves for each model were made, and the area under the curve (AUC) was computed. We evaluated the calibration of the model by drawing calibration curves. Decision curve analysis (DCA) was used to evaluate the net benefit of the prediction model for various threshold probabilities.
2.7 Statistical analyses
For statistical analysis and the creation of visualizations, GraphPad Prism 8.0, SPSS version 24.0, and R software version 4.0.2 (primarily packages for glmnet, pROC, RMS, and rmda) were utilized. The use of "glmnet" was made of the LASSO approach. The R software's "calibrate" function from the "rms" package was used for calibration. The Hosmer‒Lemeshow test was used to evaluate the calibration curve's significance. Count data were compared using the chi-square test or Fisher's exact probability test. Using the Mann‒Whitney U test, continuous variables were compared between groups. A bilateral statistical analysis was conducted, and a p value of 0.05 or lower was deemed statistically significant.