A commentary on “Multimodal radiopathomics approach for predictions of prognosis and immunotherapy response in patients with gastric cancer: a multicohort retrospective study”

Dear Editor, We read with great interest the recent publication by Wang et al titled “Multimodal radiopathomics approach for predictions of prognosis and immunotherapy response in patients with gastric cancer: a multicohort retrospective study” in the International Journal of Surgery[1]. By integrating computed tomography (CT) images with H&E stain slides features, the authors innovatively constructed a radiopathomics signature (RPS) demonstrating exceptional performance in predicting overall survival (OS) and identifying chemotherapy beneficiaries in gastric cancer (GC) patients. Their rigorous multicenter validation is particularly commendable. This study undoubtedly offers a novel approach for precision staging and treatment strategy guidance in GC. The RPS model significantly outperformed single-modality radiomics only (RS) and pathomics only (PS), in prognostic prediction, while its preliminary potential for predicting immunotherapy response also warrants attention. This study offers important insights for precision management of GC, though we propose several recommendations to enhance the biological interpretability and clinical translation potential of the RPS model. First, CT images serve as 3D macroscopic volumetric data reflecting tissue anatomy and density information, while stain slides provide microscopic 2D images revealing detailed cellular and histological morphology. Accurately aligning these two types of data poses significant challenges. The RPS model successfully integrates independently extracted CT images with H&E slide features, significantly enhancing GC prognosis and treatment response prediction. However, this paper lacks detailed descriptions on precisely performing cross-modal spatial alignment between regions acquired at different scales and dimensions. Moreover, simple feature-level concatenation appears insufficient for accurately modeling the spatial registration between medical and pathological images. The specific implementation may involve multi-step registration, such as coarse registration based on anatomical landmarks combined with non-rigid registration based on image content. Only by ensuring precise spatial alignment between the two modalities can subsequent deep fusion techniques – such as cross-modal attention networks or graph neural networks – effectively capture deep nonlinear interactions, thereby maximizing the potential of multimodal deep learning[2,3]. Second, GC treatment is inherently dynamic, whereas the RPS model primarily relies on baseline data from the initial time point. It lacks observational information on dynamic CT imaging changes and specific treatment regimens, limiting its ability to learn and analyze tumor morphological changes and the impact of different heterogeneous treatment strategies on OS. Existing research has demonstrated that dynamic features based on changes in imaging characteristics before and after treatment exhibit higher predictive value than static images[4]. In clinical practice, physicians face challenges in selecting specific treatment plans. Integrating dynamic monitoring data to provide prognostic predictions for targeted treatment strategies would significantly enhance the model’s decision-making value. Finally, the RPS model quantified and ranked the contribution of each radiopathological feature through SHAP (Shapley Additive exPlanations) analysis, which undoubtedly enhanced the model’s transparency. However, the SHAP analysis in this paper primarily focused on Shapley values for individual features, explaining the independent contribution of each feature from CT images and H&E histopathomics to prognostic prediction. It failed to delve into the complex interactions between different features, particularly cross-modal interactions. Traditional SHAP value analysis struggles to capture these “combination effects” beyond individual contributions. Calculating and analyzing SHAP Interaction Values could provide a more comprehensive understanding of how cross-modal features collectively influence model predictions, thereby enhancing the biological interpretability and credibility of RPS model decisions. In summary, the RPS model proposed by Wang’s team provides a crucial tool for GC prognosis assessment and treatment decision-making, making a substantial contribution to oncology research. We anticipate that these suggestions will motivate the authors and subsequent researchers to further explore multimodal fusion mechanisms. This approach will facilitate the development of more comprehensive and precise personalized management models for GC. This study adheres to the TITAN Guidelines 2025[5].