Radiomic model for preoperative prediction of mismatch repair deficiency in gastric cancer: a multicenter study integrating tumor sub-region radiomics and transcriptomics

Overview of the study design and research cohorts

This study was aimed at developing a model for predicting MMR status in patients with GC by using enhanced computed tomography (CT) images obtained at initial diagnosis. Moreover, we sought to integrate transcriptomic and proteomic data to elucidate the biological relevance of the radiomic model (Figure S1). A multi-center clinical and radiomic dataset comprising 1,074 patients was used (detailed information in Supplementary Methods). The patients were divided into training (dMMR, 14.8%), internal validation (dMMR, 13.2%), and external validation (dMMR, 9.4%) cohorts (Table S1) for model construction and validation. To further investigate the biological relevance of the radiomic models and their features, we performed bulk RNA sequencing on tumor samples from 55 patients with GC in the institution I cohort.

Sub-region segmentation from the region of interest

All patients underwent enhanced CT examination before radical gastrectomy. Tumor segmentation and region of interest (ROI) delineation were performed by 2 experienced radiologists in ITK-SNAP software. Radiomics feature extraction in all 4 regions included shape features, first-order features, and texture features. Square, square root, logarithm, exponential, and gradient transformation were applied to the original features (Table S2). To achieve efficient and reproducible tumor segmentation, we implemented the Simple Linear Iterative Clustering (SLIC) algorithm, an adaptive k-means clustering method specifically optimized for medical image analysis. The SLIC approach involves 3 key steps: (1) initialization of cluster centroids (seeds) distributed across the tumor volume, (2) computation of 3D gradient distances between each voxel and its neighboring seeds in both spatial (x, y, z) and radiomic (CT intensity) domains, and (3) iterative centroid updating until convergence. Through this process, each tumor was automatically partitioned into 3 phenotypically distinct sub-regions: sub-region1 (low intensity), sub-region2 (intermediate intensity), and sub-region3 (high intensity). We further conducted a comparative analysis of radiomic features across the 3 sub-regions and visualized the top 20 most differentially expressed features (Figure S2).

Radiomic panel development, model construction, and diagnostic performance

Radiomic panels associated with MMR status were developed for different spatial sub-regions of the tumors. These features were subjected to correlation analysis and LASSO dimensionality reduction, thus resulting in the selection of 32 features for the entire region, 18 for sub-region1, 29 for sub-region2, and 25 for sub-region3. The performance of each model [logistic regression (LR), support vector machines (SVM), and XGBoost] was evaluated with the area under the receiver operating characteristic curve (AUC) in each sub-region (Figure 1A). LR demonstrated higher performance than SVM and XGBoost in predicting MMR status. Among the 4 regions, the LR model for sub-region3 consistently outperformed the others in both the internal (AUC = 0.828) and external (AUC = 0.800) validation sets, by exhibiting higher AUC, accuracy, and specificity than observed in the other sub-regions (Figure 1B, Table S3).

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Radiomic models for preoperative prediction of dMMR in GC, linked with transcriptomic data. A. AUC values and 95% confidence intervals of each MMR-status prediction model constructed by 3 machine learning models in each sub-region and the entire region for each cohort. B. Radar maps of LR model performance in each sub-region and the entire region for each cohort. C. Clinical significance of models for the evaluation of dMMR risk. D. Heatmap of gene set variation analysis enrichment analyses. E. Differentially enriched functions between patients with dMMR vs. pMMR GC. F. Correlation analyses of radiomic models for MMR status prediction in each sub-region and enriched functions (P < 0.05 is represented by a dot). G. Correlation analyses between characteristic enriched functions and characteristic genes in each sub-region. H. Correlation analyses between selected radiomic features and characteristic genes in each sub-region. I. Correlation analyses between characteristic enriched functions and selected radiomic features in each sub-region. J. Correlation analyses between characteristic immune cells and selected radiomic features in each sub-region. K. IHC results of proteins coded by characteristic genes in each sub-region of patients with dMMR and pMMR. scale bar, 200 μm. L. Multiplex IHC results of the infiltration degree of characteristic immune cells in each sub-region in patients with dMMR or pMMR. scale bar, 200 μm. AUC, receiver operating characteristic curve; LR, logistic regression; SVM, support vector machines.

After the features of all 3 sub-regions were combined, the new model achieved the highest predictive performance, with an AUC of 0.863 and 0.849 in the internal and external validation cohorts, respectively (Table S3). With the optimal sub-region3 model and the combined regional model (sub-region1, sub-region2, and sub-region3), patients in the training set were stratified into low-, moderate-, and high-risk groups according to model scores. The same cutoff values were applied to stratify patients in the external validation set (Figure 1C). In the sub-region3 model, the dMMR risk rates for groups with low (< 0.3009), moderate (≥ 0.3009 and < 0.4817), and high (≥ 0.4817) risk were 6.1%, 26.1%, and 55.6%, respectively. Similarly, in the combined regional model, the corresponding risk rates were 3.7% (< 0.2067), 17.2% (≥ 0.2067 and < 0.4245), and 47.2% (≥ 0.4245). Furthermore, the calibration curves of the LR model for sub-region3 and the combined regional model, as well as the results of the decision curve analyses, all demonstrated excellent fitting performance and clinical application value (Figure S3).

Transcriptomic alterations associated with radiomic panels of sub-regions

The radiomic features selected for model construction effectively predicted the MMR status of GC, thus suggesting potential biological implications underlying these features. To explore this hypothesis, we first performed gene set variation analysis on transcriptomic data from 55 patients with GC (Figure 1D) and observed significant differences between patients with dMMR and pMMR in the enriched functions, including the TGFβ signaling pathway, Wnt/β-catenin signaling pathway, estrogen response, fatty acid metabolism, and coagulation (Figure 1E). Correlation analyses between these enriched functions with significant differences between dMMR and pMMR and the radiomic models of the 3 sub-regions indicated that the sub-region1 model was significantly associated with TGFβ signaling, whereas the models for sub-region2 and sub-region3 were associated with TGFβ signaling and coagulation, respectively (P < 0.05, Figures 1F and S4A). Further investigation identified 3,770 differentially expressed genes (DEGs) between the tumor and adjacent normal tissues in dMMR GC and 3,192 DEGs between dMMR and pMMR GC tumor tissues, with 975 overlapping DEGs (Figure S4B). Further correlation analyses between the expression of the aforementioned DEGs and the model’s scores for the corresponding sub-regions identified 22 DEGs significantly associated with the radiomic model’s score for sub-region1, 57 associated with sub-region2, and 106 associated with sub-region3 (P < 0.01, Table S4). Finally, the characteristic genes WT1 for sub-region1, NOTCH3 for sub-region2, and F12 for sub-region3 were identified according to their strong correlations with radiomic features and functional pathways (all P < 0.05, Figures 1G and S4C). To further elucidate the microscopic biological relevance of radiomic features, we conducted correlation analyses, which indicated that feature gradient_ngtdm_Busyness in sub-region1, lbp.3D.m2_firstorder_90Percentile in sub-region2, and gradient_firstorder_Minimum in sub-region3 significantly correlated with the respective characteristic genes (WT1, NOTCH3, and F12) of their respective regions (Figures 1H and S4D). Moreover, the selected radiomic feature from sub-region2 (P = 0.024) and sub-region3 (P = 0.032) demonstrated significant correlations with their corresponding characteristic functions (Figure 1I).

Immune infiltration alterations associated with radiomic panels of sub-regions

Transcriptomic data were further analyzed to evaluate the extent of immune cell infiltration within the tumor microenvironment. Correlation analyses between the infiltration degree of immune cells and the radiomic models’ scores for the 3 sub-regions identified the characteristic immune cells of M1 macrophages for sub-region1, B cells for sub-region2, and natural killer (NK) cells for sub-region3, on the basis of their high relative coefficients and significant correlations with the respective models (P < 0.05, Figure S5A). Further correlation analyses between radiomic panels and the infiltration levels of these characteristic cells revealed that wavelet.LHH_glszm_SizeZoneNonUniformity in sub-region1 and wavelet.HHL_glszm_LowGrayLevelZoneEmphasis in sub-region3 demonstrated the strongest correlations with M1 macrophages and NK cells, respectively, whereas log.sigma.1.0.mm.3D_glrlm_RunEntropy in sub-region2 showed the highest relative coefficient, with a potential but non-significant correlation with B cell infiltration (P = 0.099) (Figures 1J and S5B).

Validation of characteristic radiomic features, proteins, and immune cells

Further validation was performed on the proteins encoded by the characteristic genes and the associated immune cells identified in prior analyses. During the initial verification, we conducted targeted biopsies in each radiomics-defined sub-region from postoperative specimens. The IHC results mirrored the trends observed in the radiomic analysis. Specifically, WT1 expression in sub-region1, NOTCH3 in sub-region2, and F12 in sub-region3 were higher in the tumor tissues of patients with dMMR vs. pMMR (Figure 1K). Moreover, multicolor fluorescence in situ hybridization demonstrated that immune cells corresponding to each sub-region showed infiltration patterns consistent with the previously described trends, thus further supporting the correlations (Figure 1L).

Biological relevance and application prospects of the radiomics model and panels

The integration of radiomic models with transcriptomic data revealed distinct biological characteristics within each sub-region, which were associated with characteristic functions, genes, and immune cell infiltration associated with MMR status. For instance, WT1 in sub-region1, NOTCH3 in sub-region2, and F12 in sub-region3 emerged as potential biomarkers for identifying MMR status at specific tumor sites. Moreover, differences in immune cell infiltration, such as M1 macrophages in sub-region1, B cells in sub-region2, and NK cells in sub-region3, were observed between patients with dMMR vs. pMMR. These findings suggested that targeting characteristic immune cells might offer new opportunities to enhance immunotherapy efficacy in patients with pMMR, a subgroup typically showing poor response rates to immunotherapy. Integrating multi-omics approaches to identify novel therapeutic targets for pMMR GC is a promising research direction. However, this study had several limitations, such as the lack of spatial transcriptomic data based on spatially targeted biopsies for deeper exploration of the biological relevance across different radiomic sub-regions. This aspect is a direction for future research.

In conclusion, this study used initial CT images from patients with GC to construct and validate a radiomics model that effectively predicts MMR status through a multi-regional combined approach. By integrating spatial heterogeneity within the radiomic model with transcriptomic data, we elucidated key biological differences between patients with dMMR vs. pMMR across sub-regions. Furthermore, the characteristic functions, genes, and immune cell types identified in each sub-region provide valuable insights into the molecular and cellular bases of GC heterogeneity.

Author contributions

Guarantors of integrity for the entire study: CH, ZYX, ZSY.

Study concept/design: CH, ZYX, ZSY, SWP.

Data acquisition: all authors.

Contribution of data or analysis tools: YY, YQZ, MXC, CCH, JXX, KS.

Data analysis and interpretation: SWP, GLZ, EZL, JQZ, JY.

Manuscript drafting: SWP, EZL, GLZ, YY.

Manuscript revision for important intellectual content: CH, ZYX, ZSY, SWP.

All authors read and approved the final manuscript.