A total of 167 consecutive patients with newly diagnosed resectable LAGC who underwent gastric MRI in National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences, and Peking Union Medical College from January 2016 to December 2018 were initially enrolled. Patients were included according to the following criteria: 1) pathologically confirmed gastric adenocarcinoma on gastroscopy; 2) no contraindication to MR examinations; 3) locally advanced stage (cT3-4aN1-3M0) according to the American Joint Committee on Cancer (AJCC) TNM staging system (8th edition) (18) as determined by pretreatment CT, MRI, or endoscopic ultrasonography; 4) no previous treatment before MRI examination; 5) NCT performed within 1 week after MR examination; 6) R0 radical gastrectomy within 30 days after the completion of NCT; and 7) regular follow-up after surgery.

The enrolment flowchart of the study cohort is summarized in Figure 1. Thirty-eight patients were excluded for the following reasons: 1) low quality of MR images due to obvious respiratory movement or gastrointestinal motility artifact or significant image distortion (n = 7); 2) maximum tumor diameter (MTD) <1.0 cm (n = 5); 3) NCT not completed due to severe adverse reactions (n = 8); 4) refused surgery after complete NCT (n = 10); and 5) lost to follow-up postoperatively (n = 8). Finally, a total of 129 patients were included in this study, including 107 men and 22 women, with a median age of 60 years (range from 28 to 76 years).

Patients were asked to fast for 6–8 h prior to MR examinations to empty the gastrointestinal tract and underwent breath-holding training. In order to avoid the artifact of gastrointestinal peristalsis, patients without contraindications (i.e., glaucoma, prostate hypertrophy, asthma, or severe heart disease) were injected with 10 mg of anisodamine hydrobromide (Hangzhou Minsheng Pharmaceutical Co., Ltd., Hangzhou, China) intramuscularly, followed by drinking 800–1,000 ml of water to dilate the stomach wall before MRI.

All examinations were performed with a whole-body 3.0-T MR scanner (Discovery MR750; GE Healthcare, Milwaukee, WI, USA) equipped with an 8-channel, phased-array body coil. The conventional MRI protocols used for standardized gastric imaging at our institution include the following sequence: 1) axial three-dimensional (3D) spoiled-gradient recalled-echo sequences for liver acquisition with volume acceleration flexible (LAVA-Flex) sequence in one breath-hold; 2) respiratory-triggered axial PROPELLER T2-weighted imaging (T2WI) with fat suppression; 3) axial, coronal, and sagittal single-shot fast spin-echo T2WI in breath-hold; and 4) respiratory-triggered axial DWI sequence included two b values (b = 0 and 800 s/mm²).

IVIM-DWI was performed by using a respiratory-triggered single-shot echo-planar imaging sequence in the transverse plane with diffusion in three orthogonal directions, and the parallel imaging using the array spatial-sensitivity encoding technique (ASSET) was used to shorten the scanning time and reduce image distortion. Ten b values from 0 to 1200 s/mm² (0, 10, 20, 40, 100, 200, 400, 800, 1,000, and 1,200) were applied. Quantitative DCE-MRI was performed by using multiphase axial 3D spoiled-gradient recalled-echo sequences for liver acquisition with volume acceleration-extended volume (LAVA-XV) sequence with breath-hold. According to our previous study (19), pre-contrast T1 mapping with four different flip angles (3°, 6°, 9°, and 12°) was acquired before dynamic scanning for the determination of pre-contrast T1 values. Then a dynamic scan with 42 consecutive phases was performed, which shared the scanning parameters and range as T1 mapping, with a flip angle of 15° and temporal resolution of 6 s/phase. A bolus of gadopentetate dimeglumine (Magnevist, Bayer Schering, Berlin, Germany) at a constant dose of 0.1 mmol/kg was power injected, followed by a 20-ml saline flush at a rate of 2.5 ml/s for all patients. The acquisition time was 18 s for each of the three consecutive phases with an interval of 5–10 s; the total scanning time for DCE-MRI was 5–6 min. The detailed acquisition parameters of sequences are shown in Supplementary Table 1.

Two radiologists (YJZ and YL, with 9 and 17 years of experience in gastrointestinal abdominal imaging, respectively) who were blinded to the patients’ clinical and histopathological data independently reviewed the MR imaging and measured the DCE-MRI and IVIM-DWI parameters at the largest section of the tumor with good image quality. The mean values of quantitative parameters were used for subsequent analysis, and the interobserver agreement was also assessed according to the intraclass correlation coefficient (ICC). In order to ensure data repeatability, all parameters were measured twice with a month interval to assess the intraobserver agreement.

The region of interest (ROI) drawing principles were as follows: the slide containing the largest tumor diameter was selected for further analysis. The ROI was manually traced slightly along the borders of the tumor to include the entire tumor, while avoiding visible blood vessels, necrotic areas, and cystic areas, on DCE-MRI and IVIM-DWI.

The IVIM images were transferred to GE ADW 4.6 workstation and analyzed by MADC software in the FuncTool software package. ROI of the tumor was manually delineated on the IVIM-DWI with a b value of 800 s/mm² using axial T2WI as a reference. ADC_standard value was calculated by the monoexponential model using the total available b values according to the following equation:

The IVIM parameters were calculated by biexponential fitting according to the following equation, suggested by Le Bihan et al. (13):

where S_b is the signal intensity with diffusion gradient b and S₀ is the signal intensity without diffusion gradient. D is the true diffusion coefficient as reflected by pure water molecular diffusion, D^* is the pseudo-diffusion coefficient representing perfusion-related incoherent microcirculation, and f is the perfusion fraction related to the microvascular volume fraction. The parameter maps of IVIM were generated automatically by the MADC software, and the ADC_standard, D, D^*, and f values in the ROIs were obtained

Quantitative DCE-MRI parameters were calculated using an in-house-developed image-processing workstation, OmniKinetics 2.0.10 (GE Healthcare, Beijing, China). The signal intensity on MRI was converted into an equivalent concentration of contrast material using the variable flip angles method. The pharmacokinetic parameters including volume transfer constant (K^trans), reverse reflux rate constant (k_ep), extracellular extravascular volume fraction (v_e), and plasma volume fraction (v_p), which were derived from DCE-MRI, were calculated using the two-compartment extended Tofts model as described in our previous study and report (19, 20).

All 129 patients were treated with 4 to 6 cycles of oxaliplatin-based NCT as recommended in CSCO guideline (7), in which 76 patients receiving oxaliplatin and S-1 (SOX) regimen, 32 patients receiving capecitabine and oxaliplatin (XELOX) regimen, and 21 patients receiving docetaxel oxaliplatin and S-1 (DOS) regimen. D2 radical gastrectomy was performed within 30 days after the completion of NCT. The surgical procedures were in accordance with CSCO guidelines for gastric cancer (7). Adjuvant chemotherapy was routinely started 3–4 weeks after surgery, and the oncologist decided on both regimens and cycles based on the clinical and pathological responses.

Patient records and original histopathological slides were independently re-evaluated by 2 pathologists with over 10 years’ experience in gastrointestinal pathology. The pathologists were blinded to the routine diagnoses and patient outcomes. Response to chemotherapy was assessed according to the Mandard TRG system (21), which divided the residual tumor into grades 1–5, based on the amount of fibrosis and/or necrosis over the remaining viable tumor cells. To ensure consistency of the evaluation criteria, the 2 pathologists were trained prior to the evaluation. In case of disagreement, a consensus diagnosis would be reached through joint re-review and discussion on a multi-headed microscope. Patients were divided into two groups: pathological responders (pR) (TRG 1–3) and pathological non-responders (pNR) (TRG 4 and 5).

Histopathological variables were also recorded, including histopathological type, tumor differentiation, Lauren classification, lymphovascular invasion (LVI), perineural invasion (PNI), and immunohistochemical assays of HER2, EGFR, and c-MET. For patients with no residual remaining tumor after NCT, a preoperative biopsy specimen was used for analysis. TNM stage was assessed according to the 8th edition of the AJCC staging system. HER2, EGFR, and c-MET expression were detected according to the HER2 detection guideline from the College of American Pathologists (22) and a previous study (23).

After radical gastrectomy, all patients were followed up every 3 months for the first year and every 6–12 months afterward. Follow-up consisted of physical examination, tumor marker assessment, CT scan, and endoscopic examination. Recurrence-free survival (RFS) was recorded and was defined as the interval between the date of surgery and the first date of identified local and/or distant recurrence or the last follow-up date without recurrence. Tumor recurrence was defined as local recurrence, distant metastasis, or death caused by gastric cancer, detected by imaging or pathology. All patients were observed until recurrence or the final follow-up date of December 31, 2020. Patients were censored if they were recurrence-free and alive at the last follow-up.

All statistical analyses were performed with SPSS (version 21.0; SPSS Inc., Chicago, IL, USA). The intraobserver and interobserver reliability in the measurements of IVIM-DWI and DCE-MRI parameters was estimated with ICC, which was defined in previous studies (24).

Quantitative data were expressed as the median and interquartile range (IQR) and were compared using the Mann–Whitney U-test. Categorical data were expressed as number (percentage) and were compared using the χ² test or Fisher’s exact test, as appropriate. Differences in clinicopathological features, IVIM-DWI, and quantitative DCE-MRI parameters between the pR and pNR were compared. Those variables with a significant difference, as determined by the univariate logistic regression analysis, were chosen for multivariate logistic regression analysis to identify significant independent predictive parameters for NCT response.

Multivariate logistic regression was used to build a combined prediction model with the statistically significant parameters. The prediction performance of the quantitative parameters and models was evaluated by receiver operating characteristic (ROC) curve, quantified by the area under the curve (AUC), overall accuracy (ACC), sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV). The optimal cutoff value was calculated at the maximum value of Youden’s index (sensitivity + specificity − 1).

The quantitative parameters and prediction probabilities were converted into binary variables according to the diagnostic threshold, that is, the predicted treatment response group and non-response group. The RFS was assessed using the Kaplan–Meier method, and differences between predicted groups were assessed by log-rank test. Univariable and multivariable Cox proportional hazards analyses were used to determine independent prognostic factors for tumor recurrence among clinicopathological factors, quantitative parameters, and prediction models. p < 0.05 was considered to be statistically significant.

After the NCT and radical gastrectomy, the complete histopathologic regression of the LAGC (TRG 1) was achieved in 12/129 patients (9.3%); TRG 2 was recorded in 12/129 patients (9.3%), TRG 3 in 45/129 patients (34.9%), TRG 4 in 43/129 patients (33.3%), and TRG 5 in 17/129 (13.2%). According to TRG results, patients were divided into the pR group (n = 69) (Figure 2) and the pNR group (n = 60) (Figure 3).

There were significant differences in LVI, PNI, and postoperative pathological stage (ypStage) between the pR and pNR groups (all p < 0.001). No significant differences were found for gender, age, MTD, tumor site, surgical approach, histopathological type, differentiation, Lauren’s classification, HER2 expression, EGFR expression, and C-MET expression between the two groups (all p > 0.05) (Table 1).

The interobserver and intraobserver agreement for the assessments of quantitative parameters by the two radiologists is shown in Supplementary Table 2. ICCs for interobserver and intraobserver were all above 0.80 (95% CI, 0.837–0.975, and 0.845–0.966, respectively), which indicated excellent agreement. Therefore, the mean values of the first measurement by the two radiologists were used for subsequent analysis.

Among the 129 patients, quantitative DCE-MRI parameters in the primary lesion of K^trans, k_ep, v_e, and v_p were 0.103 (0.073, 0.161) min⁻¹, 0.592 (0.452, 0.813) min⁻¹, 0.229 (0.170, 0.351), and 0.019 (0.007, 0.037); IVIM parameters of ADC_standard, D, D^*, and f value were 1.380 (1.250, 1.650) × 10⁻³ mm²/s, 1.050 (0.920, 1.255) × 10⁻³ mm²/s, 15.400 (6.405, 34.300) × 10⁻³ mm²/s, and 41.700% (35.550%, 53.550%). The comparisons of quantitative DCE-MRI and IVIM-DWI parameters between the pR and pNR groups are summarized in Table 2. The results showed that K^trans, k_ep, and v_e values in the pR group were significantly higher than those in the pNR group (all p < 0.001), while ADC_standard and D values were significantly lower than those in the pNR group (p = 0.011 and p < 0.001). D^* value and f value in the pR group were slightly higher than those in pNR, but there were no significant differences (p-value was 0.233 and 0.105, respectively). v_p value showed no significant difference between the two groups (p = 0.470). Box and whisker plot graphs for parameters derived from DCE-MRI and IVIM-DWI in the pR and pNR groups are given in Figure 4.

With the univariate analysis, none of the clinical factors showed significant correlation with pathology response of NCT (all p > 0.05). Meanwhile, higher pretreatment baseline K^trans (OR = 9.334; 95% CI, 3.531–24.672), k_ep (OR = 4.442; 95% CI, 2.297–8.589), and v_e values (OR = 3.221; 95% CI, 1.837–5.646) and lower ADC_standard (OR = 0.673; 95% CI, 0.465–0.973) and D values (OR = 0.221; 95% CI, 0.127–0.384) were more likely to be responsive to NCT (all p < 0.05).

Multivariate logistic regression analysis demonstrated that K^trans (OR = 5.300; 95% CI, 1.470–19.104), k_ep (OR = 3.918; 95% CI, 1.484–10.345), v_e values (OR = 2.926; 95% CI, 1.437–5.961), and D values (OR = 0.266; 95% CI, 0.138–0.515) were independently associated with the response to NCT. The univariate and multivariate logistic regression results are summarized in Supplementary Table 3.

Multivariate logistic regression analysis was conducted to build prediction models for response to NCT, using K^trans, k_ep, and v_e for DCE-MRI model; D for the IVIM-DWI model; and K^trans, k_ep, v_e, and D for the DCE+IVIM model. The results are summarized in Table 3.

ROC curve analysis results of MRI quantitative parameters and combined model for predicting pathological treatment response are presented in Table 4 and Figure 5. D value was the single parameter with the highest predictive efficiency, of which AUC was 0.812. The combination of DCE-MRI and IVIM-DWI displayed the highest AUC of 0.922.

The median follow-up period for all patients was 15.0 months (IQR, 9.0–21.0 months). Of the 129 patients, 54 patients (41.9%) developed tumor recurrence by the last follow-up day. The median RFS time was 24.0 months (95% CI: 17.2–30.8), and the 2-year RFS rate was 49.9% months.

Univariate Cox proportional hazards analysis with clinicopathological factors, MRI quantitative parameters, and prediction models showed that signet ring cell, LVI, PNI, ypStage III, pathological response, K^trans value, k_ep value, v_e value, D value, DCE model, IVIM model, and DCE+IVIM model were significantly associated with RFS. Multivariate Cox regression analysis showed that ypStage III (hazard ratio [HR] = 6.197; 95% CI, 2.132–18.014) and no response predicted by IVIM model (HR = 2.240; 95% CI, 1.231–4.075) were independent predictors of poor RFS (Table 5).

The Kaplan–Meier survival analysis based on clinicopathological factors, MRI quantitative parameters, and prediction model, which were identified from Cox regression analysis, are summarized in Tables 6, 7. These features significantly stratified in 2-year RFS rate are demonstrated in Figures 6, 7 according to the log-rank test (all log-rank p < 0.05).

Further subgroup analysis showed that in ypStage II/III, LVI-positive and PNI-positive groups, and 2-year RFS rate between different groups of pathological treatment response and prediction models were also significantly different (all log-rank p < 0.05) (Table 8 and Supplementary Figure 1).