• Biomarker
• Immunotherapy
• Pathology
• Solid tumor

We read with great interest the article by Bernatowicz and colleagues.1 In recent years, immunotherapy has emerged as a groundbreaking advancement in cancer treatment, showing remarkable efficacy across diverse cancer types and significantly enhancing treatment outcomes for patients. However, the benefits of immunotherapy are limited to a subset of patients, and identifying those most likely to respond remains challenging.2 Current approaches for evaluating immunotherapy efficacy are not satisfactory. Traditional techniques, such as biopsy and histopathological analysis, are invasive, susceptible to sampling bias, and cannot reflect the dynamic evolution of the tumor immune microenvironment (TIME). Imaging modalities, though non-invasive, often lack the precision to differentiate true therapeutic responses from pseudoprogression or others. Current biomarkers, such as PD-L1 (programmed death-ligand 1) expression, tumor mutational burden, and microsatellite instability, lack universal predictive power and exhibit variability across tumor types and individuals.3 Moreover, these methods are inadequate for real-time monitoring of immune activity or capturing the spatial heterogeneity of the TIME. Thus, there is an urgent need for more accurate, non-invasive, and comprehensive tools to assess immunotherapy responses and inform personalized treatment strategies.

This study developed a radiomic method to non-invasively monitor dynamic changes in the TIME and predict immunotherapy response and patient outcome. The authors developed and validated a radiomics signature derived from routine CT data (ie, CT-TIME score), which effectively captures the spatial and temporal heterogeneity of the TIME. Using advanced machine learning algorithms, the signature demonstrated high accuracy in characterizing the intratumoral immune landscape and tracking its evolution during immunotherapy. This innovative method holds great promise for personalized cancer treatment, enabling early prediction of immunotherapy efficacy and real-time evaluation of treatment-induced changes in the TIME. As a non-invasive tool, it has the potential to revolutionize clinical decision-making and improve patient outcomes in pan-cancer immunotherapy. Despite the promising findings, we would like to highlight several concerns regarding the methodology and interpretation of the results that warrant further discussion.

First, the authors conducted a multistep radiomic feature selection process to obtain robust, non-redundant, and informative radiomic features. However, identifying radiomic features that are independent of tumor type, feature selection algorithms, and CT imaging protocols remains challenging in pan-cancer studies. It would be beneficial to perform feature selection on individual tumor types and then identify the intersection of features to discover a stable feature set consistently expressed across multiple tumors. Besides, variations in image acquisition and reconstruction settings may impact the predictive model.4 Although density normalization was performed, potential influences may still exist. Future studies should further standardize data acquisition protocols to minimize the impact of data heterogeneity on the machine learning-based radiomic model.

Second, the CT-TIME scoring system operates as a dichotomous variable, categorizing scores above 0.5 as indicative of an immune-inflamed CT-TIME status, while scores below 0.5 are classified as uninflamed CT-TIME status. However, it is unclear to the readers why 0.5 was chosen as the threshold and whether it represents the optimal cut-off. We seem unable to verify the accuracy of this threshold. Taking an extreme example, if the CT-TIME score is close to 0.5, should a score of 0.49 be classified as immune-inflamed or uninflamed CT-TIME status? It is suggested to use software such as X-tile for optimal threshold division.

Third, this study explored a meaningful connection between the CT-TIME score and underlying biological meaning across diverse cancers, enhancing the interpretability and clinical relevance of the developed model. However, the analysis of TIME biomarkers relies on biopsy samples, which may not fully capture the spatial heterogeneity of the tumor due to inherent sampling bias. This limitation is particularly pronounced in tumors with high intratumoral variability. In contrast, radiomic features extracted from CT images provide a comprehensive, global characterization of tumor heterogeneity.5 Furthermore, the CT-TIME score encompasses the characteristics of all cells, including tumor and inflammatory cells, as well as necrotic areas, while TIME-related immunohistochemical biomarkers only account for positively stained cells. Consequently, a discrepancy exists between local TIME biomarkers and the global CT-TIME score, which may contribute to the observed modest correlation coefficient. Additionally, conducting subgroup analyses based on cancer type would be highly insightful, as it may uncover the varying correlations between CT-TIME scores and TIME biomarkers, given the significant differences in TIME across different cancer types.

Fourth, the follow-up period of this study is relatively short, not exceeding 40 months. The survival tail effect of immunotherapy may impact the follow-up results, thus the follow-up time should be extended to more accurately assess the effectiveness of the model for longer prognosis, especially for tumors with relatively good prognosis.

In conclusion, this study develops an immunotherapy response signature by integrating CT-TIME radiomic signature with T cell-inflamed gene-expression profiles, providing a promising non-invasive tool for identifying potential candidates for immunotherapy and guiding clinical decision-making in patients with advanced cancer. Although this paper has some points worth discussing, it provides valuable insights for the exploration of tumor imaging biomarkers, validation of clinical effectiveness, interpretation of biological significance, and the heterogeneity of the TIME.