Editorial for “Subregional Radiomics Analysis on Multiparametric <scp>MRI</scp> for Evaluating Lymphovascular Invasion and Survival in Gastric Cancer: A Multicenter Study”

Gastric cancer (GC) remains one of the most lethal malignancies worldwide [1]. Lymphovascular invasion (LVI) is a critical risk factor associated with lymph node metastasis and poor survival outcomes [2]. Currently, the gold standard for the assessment of LVI is postoperative histopathological examination, which hinders the individualized therapeutic decision-making before surgery. Multiparametric MRI (mpMRI) serves as a potent tool for evaluation of LVI by quantifying tumor microenvironmental traits through parameters like apparent diffusion coefficient (ADC) value and Ktrans. However, the performance of these conventional features remains constrained [3]. Radiomics has emerged as a powerful tool, translating medical images into high-dimensional quantitative imaging features that reflect underlying tissue pathophysiology [4]. However, traditional radiomics approaches often focus on whole-tumor analysis, which potentially average out the spatial heterogeneity that characterizes aggressive tumor biology. In this issue of JMRI, Song et al. [5] address this limitation by presenting a multicenter study including 878 GC patients from four independent institutions. They developed and externally validated a subregion-based radiomics model for the preoperative prediction of LVI and survival in GC. They utilized Fuzzy C-Means clustering to partition tumors into two distinct subregions based on pixel intensity patterns across contrast-enhanced T1-weighted imaging, T2-weighted imaging, and ADC maps. By extracting radiomic features from these subregions and calculating the differences and ratios between them, the authors were able to quantify intra-tumoral heterogeneity. They found that the subregional radiomics model SRR_SVM significantly outperformed the traditional radiomics model based on whole-tumor analysis, achieving an AUC of 0.803 versus 0.679 in the training cohort. This superior performance was consistently maintained across both internal and external validation cohorts. Beyond binary LVI prediction, the predictive score of SRR_SVM emerged as an independent prognostic factor for overall survival and disease-free survival in multivariate Cox regression analysis. A primary limitation of this study is the lack of biological relevance for the subregions delineated by unsupervised clustering. Although Fuzzy C-Means effectively partitions tumors mathematically, the resulting subregions lack confirmed histopathological correlation. It remains unclear whether these subregions represent specific biological phenotypes, such as cellular density, necrosis, microvascular distribution, or immune microenvironment. This “black box” nature of mathematical subregions limits the interpretability of the model and hinders a mechanistic understanding of why certain subregional features predict outcomes. In contrast to radiomics studies of gliomas [6, 7], subregions are often partitioned into biologically distinct tumor core and peritumoral edema regions. Despite the need for further biological validation, the study by Song et al. [5] represents a meaningful step toward noninvasive and individualized assessment of LVI in GC. It positions subregional radiomics as a necessary evolution in the field, moving beyond the limitations of whole-tumor analysis toward a granular and subregional imaging assessment. Further studies should prioritize radiogenomic or radiopathologic correlation to bridge the gap between mathematical clustering and the actual tumor microenvironment.