Tumor-infiltrating lymphocytes

TILs represent a vital constituent of the TME and are involved in the local immune response to the TME. Previous studies have shown that the presence, differentiation and localization of TILs determine clinical outcomes, prognostic assessments and clinical decision making [38, 39]. However, specimens used for TIL assessment are not representative of TIL levels in the whole tumor due to tumor heterogeneity and differences and limitations in pathology specimen acquisition. Therefore, preoperative assessment and prediction of TIL levels by noninvasive radiomics methods are important. Table ​Table11 summarizes the studies related to radiomics prediction of TILs.

Recently, radiomics studies concerning TILs have mainly focused on breast cancer, lung cancer, and pancreatic ductal adenocarcinoma. A small number of studies have also included hepatocellular carcinoma, colorectal cancer, glioma, head and neck squamous cell carcinoma, esophageal cancer, and undifferentiated pleomorphic sarcoma. The vast majority of these studies have demonstrated good predictive performance. Radiomics studies of TILs encompass three main aspects. (1) Most studies have focused on the use of radiomics for preoperative prediction of TIL level and density. (2) A few studies have combined radiomics features with TILs to predict prognosis. (3) The TME has been defined as “cold” or “hot” based on the abundance of TILs combined with PD-L1/PD-1 expression in tumors for the prediction and exploration of relevant TME immune features.

Radiomic assessment of the TME first began with prediction of the density of a particular immune cell. Braman et al. showed that classifiers of intratumoral and peritumoral magnetic resonance imaging (MRI) radiomic features predicted the breast cancer HER2-E subtype and that features in the peritumoral 0–3 mm region correlated significantly with the density of TILs [40]. Khorrami et al. demonstrated an association between computed tomography (CT) radiomics-based peritumoral Gabor features and TIL density in response to immunotherapy in lung cancer patients [41]. Jeon et al. found a significant correlation between MRI radiomic features and CD8⁺ TIL density when predicting alterations in CD8⁺ TIL density induced by radiotherapy in locally progressive rectal cancer [42].

A study was conducted by Sun et al. to forecast the TIL level using radiomics. They effectively assessed tumor-infiltrating CD8 cells and immunotherapeutic response by CT radiomics in a variety of advanced solid tumors, and the radiomic signature of CD8 cells was validated in three additional independent cohorts, providing precise predictions for distinguishing the immunophenotypes of tumors and clinical outcomes of cancer patients undergoing anti-PD-1 and PD-L1 immunotherapy [43]. Sun et al. used validated CD8 cell radiomic features to predict the prognosis of patients with multiple cancers receiving immunotherapy in combination with radiotherapy [44], as well as to predict intercellular heterogeneity and prognosis in patients with advanced melanoma [45]. Su et al. utilized MRI radiomic features to construct a model that predicts the TIL level in triple-negative breast cancer (TNBC). Transcriptomic analyses subsequently confirmed that tumors with high TIL levels, as predicted by radiomics, exhibit activated immune-related pathways. High Rad-TIL tumors were found to possess a hot immune microenvironment characterized by upregulated gene markers for T-cell infiltration, cytokines, costimulators, and major histocompatibility complexes. Additionally, higher proportions of CD8⁺ T cells, follicular helper T cells, and memory B cells were observed in these tumors. This study demonstrated the feasibility of radiomic models to predict TIL status and provided a method by which TIL status can be interpreted [46]. In a research study on the use of MRI radiomic models to predict TIL levels in breast cancer, delayed-phase MRI radiomic features exhibited the most favorable predictive performance, with areas under the curve (AUCs) of 0.934 and 0.950, respectively [47]. Additionally, a study utilizing a CT and MRI radiomics-based extreme gradient boosting (XGBoost) classifier demonstrated that the levels of tumor-infiltrating CD8⁺ T cells in patients with pancreatic ductal adenocarcinoma (PDAC) could be effectively predicted. Furthermore, this approach can be implemented to identify potential patients who could benefit from immunotherapy [48, 49]. Liao et al. assessed tumor-infiltrating CD8⁺ T-cell levels in patients with hepatocellular carcinoma (HCC) by preoperative enhanced CT radiomics and found that Rad scores were positively correlated with the percentage of TILs [50].

A few studies have focused on combining radiomic features and TILs to predict tumor prognosis. Jimenez et al. developed a pretreatment prediction model involving the MRI radiomic signature and TIL level, a combination that improved the accuracy of predicting the pathological complete response to neoadjuvant systemic therapy in patients with TNBC [51]. Few studies have classified the TME based on PD-L1 expression and TILs, and radiomics can achieve precise prediction of the type of TME. Mazzaschi et al. demonstrated that 7 CT radiomic features could efficiently differentiate hot (inflammatory) from cold (desert) TME [52]. Trentini et al. showed that hot and cold tumor TME could be effectively distinguished by three CT radiomic features, with a predicted AUC of 0.94 for the hot tumor microenvironment [53].

In conclusion, radiomics studies of TILs are still evolving and have yielded some remarkable results, but the following limitations remain. (1) Most studies have focused on CD8⁺ T cells and have not classified TILs in a more detailed way. Therefore, more detailed studies on other types of TILs are needed to explore the detailed and comprehensive roles of TILs, especially B cells. (2) Current radiomics studies on TILs are from single centers and have small sample sizes. Therefore, a multicenter study with a large sample size is necessary to verify the robustness of the findings. (3) TILs should be combined with other important components of the TME to comprehensively reveal the characteristics of the TME for better prognosis assessment and clinical decision guidance.

Radiotranscriptomics

Radiotranscriptomics is an emerging and crucial field that investigates the correlation between radiological features derived from medical images and gene expression profiles. This research area has significant potential in cancer diagnosis, treatment planning, and prognosis assessment. Given the inherent complexity of diseases and the influence of epigenetic factors on pathogenesis, the transcriptome offers insights into sequence modifications that may not be evident at the genomic level. Transcriptomics provides a functional context for understanding the involvement of key genes and the regulatory mechanisms underlying the selective expression variation observed in disease pathogenesis. By combining transcriptomics data with imaging, a deeper understanding of the molecular intricacies of diseases can be achieved [114, 115]. Radiotranscriptomics thus emerges as a promising approach for developing noninvasive imaging biomarkers and supporting clinical decision-making.

First, numerous studies have examined the correlation between radiomics and transcriptomics, with a focus on their biological significance. For instance, Cianflone et al. conducted a study to investigate the relationship between CT imaging features and gene expression patterns in patients with ccRCC. The study successfully identified four radiogenomics patterns that exhibited a statistically significant correlation between radiogenomics features and transcriptomes [116]. Similarly, Dovrou et al. conducted a study on non-small cell lung cancer, also finding a correlation between radiogenomics features and transcriptomes. They further validated the biological relevance of these radiomic features by conducting enrichment analyses on transcriptomics-based regression models. This analysis revealed closely related biological processes and pathways, thus providing valuable radiotranscriptomics markers and models that enhance our understanding of the link between the transcriptome and phenotype in cancer [117]. Crombé et al. investigated the correlation between MRI radiomics and differential gene expression profiles in pathway analysis of soft tissue sarcomas (STSs). The study identified three reliable patient subgroups based on delta-radiomics. Group B patients exhibited increased tumor heterogeneity, necrotic signal, infiltrative margins, peritumoral edema, and peritumoral enhancement before treatment initiation. Molecular analysis revealed downregulation of natural killer cell-mediated cytotoxicity genes and upregulation of the Hedgehog and Hippo signaling pathways. This study highlights the integration of radiomics and transcriptomics in STS, suggesting that STS with extensive changes on imaging shows overactivation of suppression of proliferation and immune response [118].

Preliminary tumor molecular typing based on radiomics and transcriptomics has been addressed by several studies. Lin et al. integrated radiomics and transcriptomics analyses to provide novel insights into the molecular annotation of radiographic features and effective risk stratification in NSCLC. Three radiotranscriptomics subtypes (RTSs) were identified using radiomics and pathway enrichment profiles: RTS1 (proliferative subtype), RTS2 (metabolic subtype), and RTS3 (immune-activated subtype). RTS3 exhibited increased infiltration of the majority of immune cells. RTSs have the potential to stratify patients with molecular heterogeneity, revealing a relationship between molecular phenotypes and radiomic features [119]. In another study, Zeng et al. utilized CT radiomic features to predict gene mutations and mRNA-based molecular subtypes. They employed a combination of multiomics data (radiomics, genomics, transcriptomics, and proteomics) to predict the overall survival (OS) of patients with ccRCC, achieving a 5-year OS AUC value of 0.846. This study suggests that CT radiomics has the potential to serve as a feasible method for predicting gene mutations, molecular subtypes, and overall survival in ccRCC patients [120]. Rabasco et al. conducted a study on the molecular subtypes of locally advanced HNSCC using radiomics. They successfully improved the prognostic value of localized regional control by combining radiomic and transcriptomic features [121]. Huang et al. employed CT radiomics to differentiate between various molecular subtypes of HNSCC, including RNA-defined HPV status, five DNA methylation isoforms, four gene expression isoforms, and five somatic gene mutations. Their results demonstrated the ability of quantitative image features to differentiate between multiple molecular phenotypes [122]. In the study conducted by Le et al., transcriptome subtypes in glioblastoma patients were predicted using XGBoost-based radiomics modeling. They identified 13 radiomics features in the model that achieved high prediction accuracies for classical (70.9%), mesenchymal (73.3%), neural (88.4%), and proneural subtypes (88.4%) [123].

In recent studies, the combination of radiomics and transcriptomics has been explored. Fan et al. developed a radiotranscriptomic signature using serum miRNA levels and CT texture features to predict the response to radiotherapy in patients with NSCLC. This signature could serve as an independent biomarker for evaluating the response to radiation therapy in NSCLC patients [124]. Trivizakis et al. constructed a multiomics machine learning model that incorporated deep features and transcriptomics in an NSCLC study. The model successfully predicted molecular subtypes and histological subtypes with AUC performances of up to 0.831 and 0.925, respectively. The clinical significance of this high-performance predictive model lies in its ability to provide prognostic value and facilitate optimal therapeutic assessment [125]. Tixier et al. investigated the reflection of tumor transcriptomics through radiomics features extracted from 18F-fluorodeoxyglucose (FDG) PET images. By analyzing FDG-PET image features and transcriptomic data of patients with locally advanced head and neck cancer, they demonstrated that FDG PET radiomics features could be utilized to infer tissue gene expression and cellular pathway activity in tumors [126]. Wu et al. conducted a study that identified correlations between radiological, pathological, and molecular features of bladder cancer using multimodal data analysis, including CT, whole-slice imaging, and transcriptomics. The study identified 13 prognostically relevant features from CT and whole-slide images, and the fusion of radiologic and pathologic features achieved higher accuracy than the single-modality approach, with an AUC of 0.89. This research illustrated the potential for using multimodal data analysis in relevant clinical applications and demonstrated the correlation between CT, pathologic slides, and molecular features [127].

Despite some previous studies correlating radiomics with transcriptomics, establishing a causal relationship between the two has remained elusive. However, transcriptomics data from single cells have not yet been utilized. Thus, in future research, the integration of single-cell transcriptomics data with radiomics will lead to more precise predictions of tumor heterogeneity, microenvironments, and the efficacy and prognosis of tumor therapy.

Radiopathomics

Pathomics is an innovative interdisciplinary field that integrates the domains of pathology and artificial intelligence. It is based on the fundamental principle of utilizing a state-of-the-art full-slide scanner to digitize full-slide images, followed by the automatic extraction and classification of histological features and the subsequent conversion of this information into binary data. The extracted features are then processed through sophisticated self-learning computer algorithms, facilitating tasks such as cancer classification and outcome prediction [128]. Pathomics enables a detailed spatial exploration of the entire tumor landscape and its most aggressive elements from standard hematoxylin and eosin slides.

The development of pathomics is relatively recent, and pathomic studies are mainly focused on predicting patient prognosis. Chen et al. established that various pathomic features derived from H&E slides are independent predictors of gastric cancer prognosis and become possible potential predictors of adjuvant chemotherapy decisions [129]. Chen et al. used machine learning-based pathological features as novel prognostic markers for patients with ccRCC [130]. Fassler et al. demonstrated that pathomics can be used as an assessment of the abundance and spatial distribution patterns of TIL infiltration, an important biomarker in breast cancer, and that peritumoral and intratumoral TIL aggregation is associated with longer survival [131].

Another important research area is the combination of radiomics and pathomics. Jimenez et al. demonstrated a possible cross-scale association between digital pathology and CT imaging that could be used to identify relevant imaging and histopathological features to accurately differentiate lung adenocarcinoma from squamous cell carcinoma [132]. Brancato et al., in a glioma study, similarly showed a possible cross-scale association between digital pathomics and MRI. This not only contributes to the understanding of GBM intratumoral heterogeneity but also provides new insights into the combined multiomics approach [133]. Feng et al. developed and validated an artificially intelligent imaging pathomics fusion model using pretreatment MRI and H&E-stained biopsy slides to predict pathologic complete response in patients with locally advanced rectal cancer, with a combined model AUC of 0.812, significantly higher than other unimodal models [134]. Wan et al. effectively predicted a good pathological response after neoadjuvant radiotherapy for locally advanced rectal cancer with a model combining radiomic and pathomic features [135]. Wang et al. developed a machine learning nomogram model consisting of pathomic features, radiomic features, immune scores, and clinical features to reliably predict postoperative overall survival and disease-free survival in patients with colorectal lung metastases [60].

The concept of “digital biopsy” can be redefined through the utilization of noninvasive technology, achieved by the optimal integration of radiomics and pathomics. Additionally, the integration of radiomics and pathomics can enhance the comprehensive characterization of tumors with heterogeneity. Future studies need to prioritize the incorporation of genomics and proteomics data to construct comprehensive tumor prediction models that encompass macroscopic radiological data alongside microscopic pathological information.

Pathogenomics

Cao et al. proposed a deep learning model based on histopathology images to predict microsatellite status, and the AUCs of this model were 0.88 and 0.85. Notably, the model was able to identify five distinct pathology imaging features that were associated with mutation burden and DNA damage repair-related genotypes in the genome, as well as antitumor immune activation pathways in the transcriptome. The predictive model offers the possibility of multiomics association through interpretability associated with pathological, genomic and transcriptomic phenotypes [136]. Ziemys et al. demonstrated the potential of digital pathomics in transcriptional analysis combined with the spatial distribution of immune cells to effectively predict the clinical response to BRAF inhibition in metastatic melanoma [137]. Huang et al. identified N7-methylguanosine modification (m7G) in pancancer as an innovative marker for predicting clinical outcome and immunotherapy efficacy. Cellular pathway enrichment analysis showed that m7G scores were strongly associated with invasion, the cell cycle, and DNA damage and repair. In several cancers, m7G scores were positively correlated with tumor immune dysfunction and exclusion. The XGBoost-based pathomics model accurately predicted m7G scores with an AUC of 0.97. Furthermore, scRNA-seq analysis showed that m7G significantly differed between cells in the TME, indicating its potential as a therapeutic target for cancer treatment [138]. Yu et al. reported good performance in predicting survival and immunotherapy response in cervical cancer patients by pathomic characterization and genomic profiling [139]. Sammut et al. integrated digital pathology and genomic and transcriptomic data from pretreatment breast cancer samples and correlated the pathological outcomes of surgery (complete remission or residual disease) with a multiomics profile. Research has demonstrated that the response to treatment is influenced by the pretreatment tumor ecosystem, and its multiomics landscape can be incorporated into predictive models using machine learning. The degree of residual disease after treatment is linearly related to pretreatment characteristics, such as tumor mutation and copy number landscapes, tumor proliferation, immune impact, and T-cell dysfunction and rejection [140].

Pathomics, as an emerging research method, is currently in the early stages of investigation. Despite the significant potential and prospects of supplementing and enhancing histopathology with other omics data, current research is primarily limited to small-scale datasets from single institutions. Subsequent studies that offer large-scale, multimodal datasets may facilitate the development of elaborate integration strategies, further unleashing the potential of pathomics and quantitative tissue morphology.

Thanks for the drawing platform of “Figdraw”, Fig. ​Fig.11 was drawn with “Figdraw”.