CT-based radiomics nomograms for preoperative prediction of diffuse-type and signet ring cell gastric cancer: a multicenter development and validation cohort

This retrospective study was approved by the institutional review board of two medical centers, and the need for informed patient consent was waived. All procedures performed involving human participants were following the 1964 Helsinki Declaration and its later amendments. Patients who underwent total or partial radical gastrectomy and histologically confirmed GC between December 2007 and March 2016 were enrolled. The detailed inclusion criteria were as follows: (1) patients who underwent surgery for GC; (2) patients who underwent standard contrast-enhanced CT less than 15 days before surgery; (3) patients with complete clinicopathologic data. Patients who received neoadjuvant chemotherapy (NAC) therapy or radiotherapy before surgery were excluded to avoid the influence of these factors on the tumor size and degree of invasion. The demographic and clinicopathologic data of patients, including age, sex, tumor site, tumor size (maximum diameter), CEA, CA199, Lauren classification, Borrmann classification, differentiation and tumor stage, were obtained from medical records. Tumor staging was performed based on the American Joint Committee on Cancer tumor-node-metastasis (TNM) Staging Manual, 8th Edition.

Flow diagrams for eligible patients were shown in Additional file 1: Figure S1. Finally, a total of 693 patients (453 males and 240 females; mean age, 56.38 ± 11.85 years; age range, 22–87 years) from 2 medical centers were enrolled in the study, including 587 patients from center 1 (Nanfang Hospital of Southern Medical University, Guangzhou, China) and 106 patients from center 2 (Zhujiang Hospital, Guangzhou, China). To develop a Lauren radiomics model to identify diffuse-type GC, we divided all patients into three cohorts: one training cohort (n = 300 from center 1), one internal validation cohort (n = 287 from center 1) and one external validation cohort (n = 106 from center 2) (Additional file 1: Figure S1a). Moreover, the SRCC radiomics model was designed to identify SRCC from diffuse-type GC. A total of 443 diffuse-type GC patients were included and divided into three cohorts: one training cohort (n = 280 from center 1), one internal validation cohort (n = 114 from center 1) and one external validation cohort (n = 49 from center 2) (Additional file 1: Figure S1b). The sample size consideration was shown in Additional file 1: S1.

The procedures of CT image acquisition and retrieval were described in detail in Additional file 1: S2. Then CT images were exported to the ITK-SNAP 3.6 (ITK-SNAP 3.X TEAM) software, and three-dimensional (3D) segmentation of the region of interest (ROI) was performed (Additional file 1: Figure S2). The algorithms for tumor ROIs delineation and reproducibility evaluation of intraobserver (reader 1 twice) and interobserver (reader 1 vs. reader 2) were described in Additional file 1: S3. The preprocessing was applied to the ROI images with different parameters (Additional file 1: Table S1) to enrich features before extracting the texture features (Additional file 1: S4). Then we applied the feature extraction method to the ROI in MATLAB 2016b (Mathworks), and series of texture features were generated from the images (Additional file 1: Table S2). Then the feature values were preprocessed with a filtering feature selection method (Additional file 1: S5).

We used the Relief forward selection (RFS) algorithm [25] and an exhaustive test based on the performance of the SVM classifier (Additional file 1: S6) to find feature subset with the best distinguishing characteristics for the radiomics model [25]. The SVM model was based on the LIBSVM software package developed by Professor Lin et al. in 2001 (https://​www.​csie.​ntu.​edu.​tw/​~cjlin). A high penalty parameter c could effectively improve the model’s accuracy, but an excessive high penalty parameter would cause over-fitting status. The range of c was limited to prevent this situation, and the tenfold cross-validation and grid search method was applied to find the best combination of SVM model parameters (c and g) (Additional file 1: Figure S3). Then all extracted features were ranked from the most important to the least important, and different feature sets were obtained using the exhaustive test from the ordered sequence 1 ≤ m ≤ M. The set of first m features was fed into the SVM classifier. Its performance for differentiating different GC types was evaluated by receiver operating characteristic (ROC) curves and the area under the curve (AUC). The detailed steps of feature selection were shown in Additional file 1: S7 in the Supplement. Differences in the AUC values between the three cohorts were assessed using the Delong test. The pathological classification radiomics feature score obtained in SVM models of each patient was seen as Rad-score. Kaplan–Meier survival analyses were used to estimate the difference in 5-year disease-free survival (DFS) and 5-year overall survival (OS) between the high Rad-score and low Rad-score groups.

Multivariate logistic regression was applied to select independent predictors of diffuse-type and SRCC GC from the clinical characteristics. The significant predictors among the clinical characteristics and the Rad-score were entered into the logistic regression analysis to develop the radiomics nomogram. The diagnostic performance and calibration of the radiomics nomogram were evaluated based on ROC and calibration curves. Decision curve analysis was applied to assess the clinical usefulness of the radiomics nomograms by quantifying the net benefit at different threshold probabilities.

A total of 693 patients (453 males and 240 females; mean age, 56.38 ± 11.85 years; age range, 22–87 years) were included in the study. The clinicopathologic characteristics of the assessed patients were listed in Table 1. Clinical characteristics, including tumor location, differentiation status, Borrmann type, levels of CEA and CA199, and TNM stages were significantly different between intestinal-type and diffuse-type GC patients.

A total of 9691 features were extracted from the tumor ROI with satisfactory interobserver and intraobserver reproducibility assessments (Additional file 1: S8). The weight ordering of radiomics features was obtained by the Relief algorithm (Fig. 1a). The feature subset with the best discrimination ability for the radiomics model was obtained using the exhaustive test based on the performance of the support vector machine (SVM) classifier. Finally, the optimal feature subset with 13 features achieved excellent performance in distinguishing Lauren diffuse-type and intestinal-type GC (Fig. 1b and Additional file 1: Table S3), yielding AUC values of 0.895 (95% confidence interval (CI) 0.957–0.932), 0.791 (95%CI0.728–0.853) and 0.857 (95%CI0.78–0.935) in the training, internal validation and external validation cohort, respectively (Fig. 1c). Multivariate analysis revealed that Rad-score was the significant predictor between intestinal-type and diffuse-type GC (OR, 4.164; 95%CI, (3.121,5.557); p < 0.001) (Additional file 1: Table S4). Further, statistical difference was found in terms of 5-year DFS and OS between the high Rad-score group (diffuse-type) and low Rad-score group (intestinal-type) (Additional file 1: S9).

Univariate logistic regression analyses showed that age, tumor size, tumor location, and elevated CEA and CA199 had statistically significant p-values between the diffuse-type and intestinal-type GC patients. Multivariate logistic analysis revealed that age (OR, 0.979 [0.958, 1.000], p = 0.049), tumor location (OR, 1.347 [1.035, 1.753], p = 0.027) and elevated CEA (OR, 2.302 [1.417, 3.740], p = 0.001) were independent predictors (Table 2). The Rad-score and clinical characteristics were incorporated to build the Lauren radiomics nomogram (Fig. 2).

The diagnostic performance comparison of the Lauren radiomics model and radiomics nomogram was shown in Fig. 3. No difference was observed in the training cohort (Fig. 3a), while the radiomics nomogram achieved a higher AUC than the radiomics model in the internal validation cohort (AUC, 0.841 [95%CI0.781–0.901] vs 0.791 [95%CI 0.728–0.853]) and external validation cohort (AUC, 0.893 [95%CI 0.831–0.955] vs 0.857 [95%CI 0.78–0.935]) (Fig. 3b and 3c). The Lauren radiomics nomogram model had higher specificity, sensitivity and accuracy than the SVM model (Additional file 1: Table S5). The Delong test was applied on the ROC curves of the radiomics nomogram to assess possible overfitting and the result revealed that the differences were not statistically significant among the AUCs of the training cohort and the two validation cohorts, with P values of 0.138 and 0.969, respectively. The calibration curves demonstrated good agreement between prediction and observation in all three cohorts (Hosmer–Lemeshow test, p > 0.05) (Fig. 3d–f). The decision curve analysis (Additional file 1: Figure S6) indicated that the patients would benefit more from using the radiomics nomograms than using the SVM model or treat-all-patients scheme or the treat-none scheme if the threshold probability in the clinical decision was between 10 and 90%.

To further develop the SRCC radiomics model to distinguish SRCC and non-SRCC in diffuse-type GC patients, 394 diffuse-type GC patients from center 1 and 49 patients from center 2 were enrolled. The clinicopathologic characteristics of these patients were listed in Table 3. Clinical characteristics, including tumor location, differentiation status, Borrmann type, levels of CEA and CA199, and TNM stages were significantly different between non-SRCC and SRCC GC patients.

Following the same feature selection and SVM model building procedure, we searched the feature subset with the best distinguishing characteristics for the SRCC radiomics model. Finally, the optimal feature subset with 10 features achieved excellent performance in distinguishing SRCC and non-SRCC patients (Fig. 4a and Additional file 1: Table S6), yielding AUC values of 0.904 (95%CI0.865–0.942), 0.824 (95%CI 0.748–0.9) and 0.835 (95%CI 0.709–0.962) in the training, internal and external validation cohort, respectively (Fig. 4b). Multivariate analysis revealed that Rad-score was the significant predictor of SRCC (OR, 6.193; 95%CI, (4.123, 9.303); p < 0.001) (Additional file 1: Table S4). Further, statistical difference was found in terms of 5-year DFS and OS between the high-Rad-score (SRCC) and low-Rad-score (non-SRCC) groups (Additional file 1: S9).

Multivariate logistic regression analyses revealed that sex (OR, 2.044 [1.228, 3.404], p = 0.006) and tumor location (OR, 1.449 [1.058, 1.984], p = 0.021) were independent predictors between the SRCC and non-SRCC patients (Table 3). The Rad-score and clinical characteristics were incorporated to build the SRCC radiomics nomograms (Fig. 5).

The diagnostic performance comparison of the SRCC radiomics model and radiomics nomogram was shown in Fig. 6. No obvious differences were observed in the training cohort (Fig. 6a), while the radiomics nomogram achieved higher AUC than the radiomics model in the internal validation cohort (AUC, 0.845 [95%CI 0.775–0.915] vs 0.824 [95%CI 0.748–0.900]) (Fig. 6b) and external validation cohort (AUC, 0.918 [95%CI 0.842–0.994] vs 0.835 [95%CI 0.709–0.962]) (Fig. 6c). The SRCC radiomics nomogram model had higher specificity, sensitivity and accuracy than the SVM model (Additional file 1: Table S7). The Delong test revealed that the differences were not statistically significant among the AUCs of the training cohort and the two validation cohorts, with P values of 0.138 and 0.969, indicating no overfitting was assessed. The calibration curves of the radiomics nomogram demonstrated good agreement between prediction and observation in all three cohorts (Hosmer–Lemeshow test, p > 0.05) (Fig. 6d–f). The decision curve analysis indicated that the patients would benefit more from using the radiomics nomograms than using the SVM model or treat-all-patients scheme or the treat-none scheme if the threshold probability in the clinical decision was between 10 and 90% (Additional file 1: S9).