A commentary on “Multimodal radiopathomics approach for predictions of prognosis and immunotherapy response in patients with gastric cancer: a multicohort retrospective study”
Year: 2026
Type: article
Abstract: To the editor, We read with interest the article by Wang et al, who developed and validated a multimodal radiopathomics signature for predicting prognosis and treatment response in gastric cancer. The integration of computed tomography (CT) radiomics with digital histopathology features represents a significant step toward more personalized oncology[1]. However, we wish to elaborate on two persistent challenges inherent to such multimodal artificial intelligence studies, which the article underscores but cannot fully resolve. First, the technical intricacies of fusing fundamentally disparate data modalities remain formidable. While the authors adeptly employed deep learning architectures for feature extraction, the core challenge lies not in extraction but in integration. CT images and hematoxylin and eosin stained whole slide images exist on different scales, capture orthogonal biological information, and are susceptible to distinct preprocessing artifacts. The process of creating a unified feature space from such heterogeneous inputs risks information loss or the introduction of spurious correlations[2]. The study used LASSO regression for feature selection, but it remains unclear whether this linear combination truly captures the synergistic, potentially nonlinear, interactions between macroscopic imaging phenotypes and microscopic tissue architecture. Such methodological considerations are crucial, as oversimplified approaches can be a source of failure in medical imaging AI[3]. Second, the compelling performance in external validation cohorts is promising, yet it invites cautious optimism. Variations in clinical protocols can introduce “center effects” that silently degrade model performance, a recognized hurdle in evaluating clinical AI[4]. The authors acknowledge the need for prospective validation. However, can retrospective harmonization of imaging protocols fully compensate for the underlying biological and technical heterogeneity present in real world, global clinical practice? This underscores the imperative for validation frameworks that meet stringent reporting standards for clinical AI modeling[5]. The model’s performance in the immunotherapy cohort, while suggestive, is based on a limited sample size. Demonstrating robustness across this spectrum of clinical variance is the next necessary hurdle for delivering tangible clinical impact[6]. In conclusion, the work provides a valuable proof of concept for radiopathomics in gastric cancer. To translate such models from robust retrospective analysis to reliable clinical tools, the community must prioritize solutions for intrinsic data fusion challenges and commit to stringent, prospective, and multinational validation efforts. Ethical approval Not applicable. Consent Not applicable.
Source: International Journal of Surgery
Authors: Aibibai Mohetaer, Diliyaer Adili
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
Estimated APC paid: $3,450
OpenAlex ID: W7124241482