Radiomics-guided checkpoint inhibitor immunotherapy for precision medicine in cancer: A review for clinicians

Immunotherapy using immune checkpoint inhibitors (ICIs) is a breakthrough in oncology development and has been applied to multiple solid tumors. However, unlike traditional cancer treatment approaches, immune checkpoint inhibitors (ICIs) initiate indirect cytotoxicity by generating inflammation, which causes enlargement of the lesion in some cases. Therefore, rather than declaring progressive disease (PD) immediately, confirmation upon follow-up radiological evaluation after four–eight weeks is suggested according to immune-related Response Evaluation Criteria in Solid Tumors (ir-RECIST). Given the difficulty for clinicians to immediately distinguish pseudoprogression from true disease progression, we need novel tools to assist in this field. Radiomics, an innovative data analysis technique that quantifies tumor characteristics through high-throughput extraction of quantitative features from images, can enable the detection of additional information from early imaging. This review will summarize the recent advances in radiomics concerning immunotherapy. Notably, we will discuss the potential of applying radiomics to differentiate pseudoprogression from PD to avoid condition exacerbation during confirmatory periods. We also review the applications of radiomics in hyperprogression, immune-related biomarkers, efficacy, and immune-related adverse events (irAEs). We found that radiomics has shown promising results in precision cancer immunotherapy with early detection in noninvasive ways.


Introduction
Immunotherapy using ICIs has been revolutionary in cancer treatment owing to its significant impact on the reactivation of the immune system (1,2). Unlike traditional cancer treatment approaches, which kill tumor cells directly, ICIs initiate indirect cytotoxicity by generating inflammation and may cause enlargement of the lesion in some cases. Hence, there may be different interpretations of medical imaging for patients undergoing immunotherapy (3).
Medical images contain many quantitative biomedical features based on intensity, shape, size or volume, and texture, which can offer information on the tumor microenvironment and phenotype. These features are difficult to identify by human vision alone.
Radiomics is an emerging field that extracts quantitative features from medical images and converts digital medical images into mineable, high-dimensional data with new highthroughput approaches.
The process of radiomics involves the following discrete steps: 1. Image acquisition (i.e., CT, MR, and PET/CT) 2. Volume of interest (VOI) identification and segmentation: identifying tumors and their surroundings as VOIs and delineating the borders of the volume 3. Feature extraction and qualification: extracting and qualifying high-dimensional features from the VOI 4. Modeling: mining extracted features with artificial intelligence to develop classifier models that aid detection, diagnosis, prognosis assessment, and treatment response prediction This approach could be applied to any aspect of medical imaging analysis, including immunotherapy, thereby providing a novel noninvasive approach to precision cancer treatment.
Genomic and microenvironment heterogeneity within the tumor volume is displayed on the imaging, while these fine distinctions cannot be recognizable by the naked eye, even for experienced radiologists. Nevertheless, these subtle differences can be recognized by radiomics using quantitative assays, allowing for microscopic analyses of medical imaging to establish predictive, diagnostic, and prognostic models ( Figure 1).
This review summarizes radiomics concerning immunotherapy from a clinical perspective. It discusses its potential to predict outcomes, molecular biomarkers, atypical responses, and immune-related adverse events (irAEs) of immunotherapy.

Prediction of immune-related biomarkers
To date, predictive biomarkers of immune responses are mainly driven by invasive tissue biopsy, while limited biopsy samples may be difficult to provide a holistic picture of the heterogeneity within the tumor and its microenvironment. Radiomics is a powerful auxiliary FIGURE 1 Radiomics analysis can obtain more information from medical images. (A) There is genomic and microenvironment heterogeneity within the tumor volume displayed on the imaging, while these fine distinctions cannot be recognizable by the naked eye, even for experienced radiologists. (B) These subtle differences can be recognized by radiomics using quantitative assays, allowing microscopic analyses of medical imaging to establish predictive, diagnostic, and prognostic models.
to conventional invasive biopsies, overcoming the intratumoral heterogeneity within the same patient. The section below describes current progress in immune-related biomarkers using radiomics and how radiomics can overcome these limitations (Table 1).

Radiomics and programmed cell death ligand 1
As a promising treatment for cancer patients, immunotherapy is not effective for all patients (3,29,30). Thus, recognizing the appropriate candidate for immunotherapy is of vital importance. PD-L1 expression examined via immunohistochemistry (IHC) is associated with the clinical efficacy of anti-PD-1/PD-L1 therapy. It has been widely applied as a reference to immunotherapy decision-making in most cancer types during clinical practice (29,31). The expression levels of PD-L1 may change during therapy (32,33). Due to the existence of intratumoral heterogeneity, the IHC test results of a small number of biopsy samples could not be representative of the whole (34)(35)(36).
As the artificial intelligence (AI) field progresses, PD-L1 prediction with radiomics has received increased attention in recent years (Table 1). Giulia Mazzaschi et al. established a noninvasive model with computed tomography (CT)-extracted features to predict the level of PD-L1 expression and tumor infiltrating lymphocytes (TILs). They found that texture, effect, and margins were directly associated with PD-L1 expression and TILs. These features also correlated with the prognosis of patients with non-small cell lung cancer (NSCLC) (5). Similarly, researchers have tried to explore the potential of radiomics biomarkers for predicting the immunooncologic characteristics of hepatocellular carcinomas (HCC) based on magnetic resonance imaging (MRI). They found that radiomics features, specifically texture feature variance and enhancement ratios, were strongly associated with PD-L1 expression and were predictive diagnostic biomarkers for assessing early HCC recurrence (6).
PET/CT was also applied to predict PD-L1 expression levels. The researchers extracted imaging histology features from PET/CT images of 399 lung cancer patients, of which 24 features were closely related to PD-L1 expression levels. The researchers further developed prediction models based on these features. PET/CTbased prediction model achieves 88% AUC in predicting patients with >50% PD-L1 expression (8).
Radiomics features, in combination with clinical characteristics, may present a better predictive performance. Yoon et al. built a PD-L1 prediction model using both the Rad-score and clinical variables, which turned out to be more accurate than the clinical-variableonly derived model (7).
To date, radiomics remains insufficient to replace IHC testing for the detection of PD-L1. Despite this, the repeatable and noninvasive nature of radiomics analysis may offer additional information for difficult-to-repeat invasive PD-L1 analyses.

Radiomics and TILs
TILs in the tumor microenvironment play an essential part in the immune response against cancer and appear to be associated with the outcome of immunotherapy. Previous studies have indicated that tumor-infiltrating regulatory T cells (Treg) and t u m o r -a s s o c i a t e d m a c r o p h a g e s ( T A M s ) i n d u c e a n immunosuppressive microenvironment that is directly responsible for the failure of immunotherapy, while CD8+ T tumor infiltration is associated with better outcomes to cancer immunotherapy (37).
One of the early attempts to create radiomics signatures aimed at predicting the presence of CD8+ T cells and the clinical efficacy of immunotherapy was conducted by Sun et al. The signature combining eight radiomics features was developed from CT images, genomic data, and ribonucleic acid (RNA) sequencing. The signature was validated on external cohorts to discriminate CD8+ cells and immune phenotypes. Researchers also found that a higher baseline radiomics score is associated with a better response to immunotherapy (12). This study used a radiomics-based biomarker to correlate pathology with prognosis, while the area under the ROC curve (AUC) score for this prediction model was relatively low. In addition, the parameters of image acquisition were not uniform. Therefore, the credibility of the radiomics signature would be affected (12). Another retrospective study revealed that low CT image intensity and high heterogeneity were associated with lower PD-L1 expression and higher CD3 cell infiltration, which was an immune-activated state strongly correlated with favorable overall survival (OS) (13).
Many studies extract radiomics features from pretreatment or posttreatment medical images to predict the TILs associated with the response to immunotherapy (14,15). "Delta radiomics" can explore changes in the tumor microenvironment before and after immunotherapy. Therefore, it may provide an earlier and more accurate prediction of the efficacy of immunotherapy before visible changes to the naked eye. Khorrami and colleagues explored the potential of radiomics using pretreatment and subsequent CT images from 135 NSCLC patients treated with ICIs. The concordance of the radiomics features with TIL infiltration was confirmed by comparison with TIL infiltration in diagnostic biopsy samples. They reported that delta-radiomics is associated with response and OS in NSCLC patients undergoing ICIs (16).

Radiomics and tumor mutational burden
Previous studies have suggested that TMB is another predictive biomarker for immunotherapy across multiple cancer types, as high TMB is correlated with greater neoantigen and immune infiltration (38).
Researchers have investigated the potential of applying radiomics to predict the TMB status in patients with advanced NSCLC and endometrial and bladder cancers (17, 18,   Radiomics has the potential value of being a powerful aid in classifying TMB status. Specifically, combining clinical and pathological features may improve prediction performance (14)(15)(16). Radiomics may provide sufficient information despite intratumor heterogeneity to assist clinical decisions on immunotherapy.  Table 1).

Radiomics and other predictive biomarkers of immunotherapy
CAO et al. evaluated whether CT-based radiomics can predict MSI status in 502 patients with colorectal cancer (CRC) based on preoperative contrast-enhanced CT images. They further combined the clinical characteristics with radiomics and then developed a nomogram to predict the MSI status. The predictive performance of the radiomics-clinical nomogram was superior to radiomics only and clinical only (24). Such findings are consistent with another study (25). Combining clinical and pathological features with radiomics may add specificity to the prediction model and contribute to personalized clinical decision-making. The MRIbased MSI prediction model was similarly developed, with an AUC of 0.868 (26). Jiang et al. immunoscored gastric cancer patients based on the immunohistochemical expression of CD3, CD8, CD45 and CD66b and classified the immunoscores using CTbased radiomics model. The radiomics model accurately classified patients with high immunoscore and low immunoscore with an AUC of 0.786 and had the potential to select patients who would benefit from chemotherapy (27).
Predictive biomarkers of immunotherapy, including PD-L1 and TMB status, are acquired via biopsy, which is invasive, difficult to perform dynamically, and restricted to a small sample of pathological specimens.
Radiomics offers a noninvasive whole-body evaluation of tissue biomarkers. Heterogeneity within the tumor may harbor prognostic information that can be captured and transferred into radiomics features by radiomics analysis. Considerable evidence suggests that radiomics could predict immunotherapy efficacy by recognizing radiomics features associated with response (Table 2).
Trebeschi et al. developed a model able to predict the response to immunotherapy. They suggested that greater heterogeneous and nonuniform lesions were associated with a better response in NSCLC, possibly with infiltration and inflammation of the tumor, while the sample size of the melanoma cohort was too small to identify optimal imaging biomarkers (54).
Skewness, representing the heterogeneity of a segmented lesion, was a significant independent predictor of OS and PFS, with a higher skewness value linked with poorer survival (55). Similarly, Velcheti analyzed the CT features of 50 NSCLC patients who underwent nivolumab, and they found that vessel tortuosity was an independent predictor of nivolumab's efficacy (56).
Radiomics features from 18F-FDG PET/CT scans also showed the ability to predict the response to immunotherapy. Several texture features (PET_SRLGE, KLD_SZE) were associated with durable clinical benefits, demonstrating that patients with more heterogeneous tumors might benefit more from immunotherapy. Notably, the prospective cohort validated the model with an AUC of 0.81 (57). However, these results are somewhat inconsistent with the results from prior studies, which indicated that more heterogeneous tumors with CT textures are associated with worse response rates to chemotherapy or radiation.
The predictive performance of the radiomics model (consisting of small run emphasis and difference entropy) developed by Valentinuzzi et al. was superior to that of PD-L1 and iRECIST, with an AUC of 0.90. Specifically, small run emphasis has the highest predictive performance to discriminate survival, with higher small run emphasis possibly having OS survival from pembrolizumab treatment (63). These results reflected that patients with more homogeneous tumors might benefit from immunotherapy, consistent with a previous study conducted by Polverari et al., where nonresponders exhibited higher tumor heterogeneity at pretreatment CT images (reflected by higher kurtosis and skewness) than responders (11). A study conducted by Mu et al. showed the opposite result. They found that heterogeneous tumors might be more likely to achieve durable clinical responses (57).
These studies have demonstrated that radiomics has the potential to predict the response to immunotherapy and could facilitate clinical decision-making. Despite its great potential, the application of radiomics in clinical immunotherapy is still in its infancy. Reproducibility and standardization are major problems. Studies have already explored the standardized workflow of radiomics (70)(71)(72)(73)(74).

Atypical response
In the widespread use of immunotherapy in cancer treatment, unconventional characteristics of response, so-called atypical   (78)(79)(80). In this part, we summarize how radiomics can support clinical decision-making in light of atypical responses (Table 3).

Pseudoprogression
Pseudoprogression is defined as an initial radiographic increase in tumor size or the appearance of new lesions, followed by a response (76,87,88). This radiologic effect is due to inflammatory cell infiltration around tumor cells, with an apparent increase in size, and can be confused with tumor cell proliferation (79,87,89). To discern pseudoprogression, immune-related response criteria were developed, which suggest further radiological evaluation after four to eight weeks rather than declaring PD immediately (90). A switch to next-line treatment might be delayed while waiting for the confirmatory follow-up evaluation. It is crucial to distinguish pseudoprogression from true disease progression in a timely manner, as it is highly relevant in daily clinical decision-making processes. Therefore, early detection of pseudoprogression is of vital importance.
Texture features extracted from radiological images have been identified to distinguish inflammatory infiltration from tumor cell proliferation. Basler et al. evaluated the capability of PET/CT-based radiomics features, lesion volume, and routine blood markers to differentiate pseudoprogression from true progression at the third month. Of the seven models constructed based on blood, volume, and radiomics, the blood-radiomics model has the best predictive performance, achieving the highest AUC (0.82), and it is a promising biomarker for the early differentiation of pseudoprogression in the third month (81). In a single-center retrospective study, Ji et al. used four radiomics models constructed using contrast-enhanced CT of 87 patients with lung cancer treated with ICIs; of these, model three and model four had AUCs of 0.736 and 0.760, respectively, and were both accurate in predicting responses in two of three pseudoprogression patients (59). Similarly, He et al. used a radiomics approach to identify pseudoprogression from true progression. They extracted intratumoral and peritumoral radiomics features from baseline chest CT scans of 135 patients and built a predictive model. The model had an AUC of 0.96 in the validation set (82). These findings suggest that radiomics can predict pseudoprogression in the course of immunotherapy and may supplement immune-related response evaluation criteria.

Hyperprogression
Hyperprogression is an atypical response to ICIs with paradoxical tumor growth acceleration soon after immunotherapy initiation (91-94). It has been described in numerous cancer types with an incidence of 6%-29% (95,96). It has been associated with high metastatic burden, significantly shortened survival, and poorer performance status (95, 97-100), thus limiting the potential for administration of other therapies. Identifying high-risk groups is vital, yet there are no predictive biomarkers with apparent effects to identify the risk of hyperprogression (96,98). As a noninvasive method, radiomics has been explored for the risk stratification of hyperprogression in patients undergoing immunotherapy.
To evaluate the accuracy of a pretreatment CT-based radiomics model in predicting hyperprogression in NSCLC patients treated with ICIs, Vaidya et al. found that peritumoural texture and vasculature patterns in the baseline CT scans positively correlate with hyperprogression. These features had higher expression in hyperprogression than in responders or nonresponders, suggesting that patients with more heterogeneous tumors are more likely to derive FIGURE 2 Atypical patterns in patients undergoing immunotherapy. (A) Hyperprogression is an atypical response to ICIs with a paradoxical acceleration of tumor growth soon after the initiation of immunotherapy. (B) Dissociated Response is defined as a reduction at baseline or increase < 20% in target lesions compared with a nadir in the presence of the new lesion. (C) Pseudoprogression is defined as an initial radiographic increase in tumor size or the appearance of new lesions, followed by a response.  (82). PET/CT-based radiomics has also been attempted for predicting hyperprogression. Gabryśet al. developed a predictive model using PET/CT of patients with metastatic melanoma. CT-based radiological features were shown to be better predictors of hyperprogression than PET-based features (85). Given the poor prognosis of hyperprogression, it is of great importance for high-risk populations to be screened before initiating immunotherapy. The remarkably accelerated development of radiomics in immunotherapy suggests that radiomics could be used to stratify the risk of hyperprogression. These findings warrant further exploration.

Radiomics and irAEs
IrAEs associated with immunotherapy, resulting from activating an immune response against healthy tissues, may involve almost every organ and system (101,102). Timely diagnosis and prompt management depending on its severity, with a proper suspension of ICIs or corticosteroid treatment, is vital. If left untreated, irAEs could develop into life-threatening complications. Therefore, early diagnosis and monitoring of irAEs are crucial for radiologists.
Medical imaging, including CT, ultrasonography, magnetic resonance imaging, X-rays, and PET/CT, was used to detect irAEs. In a retrospective study with a small sample size, Mekki et al. found that 74% (19) of irAE patients (55) showed abnormalities on medical imaging and could be diagnosed by radiologists. The rates of enterocolitis, hypophysitis, thyroiditis, hepatitis, arthralgia or arthritis, lung/mediastinum side effects, and pancreas range from 28% to 100% (103).
The detection of irAEs generally depends on blood test indicators, clinical manifestations, and imaging characteristics. However, radiomics can help in the identification of early invisible signs of irAEs in medical imaging. Colen et al. utilized a radiomics approach to predict the risk for immune-related pneumonitis. They extracted 1860 radiomics features from baseline chest CT scans of 32 patients treated with ICIs, of whom two developed immune-related pneumonitis. Selective radiomics features were utilized to develop the predictive model of the subsequent development of pneumonitis. This model correctly identified the two patients who developed immune-related pneumonitis (86). Despite the noninvasive and impersonal nature of radiomics, studies relevant to the early detection of irAEs are rarely reported (Table 3).

Discussion
With the rapid development and application of artificial intelligence in medicine, radiomics may become a valuable tool in clinical decision-making. Here, we focus on the exciting and innovative space of radiomics to solve the problems in immunotherapy and discuss how radiomics serves as a means to support the precision design of immunotherapy, especially in pseudoprogression and hyperprogression.
Radiomics represents a potential noninvasive and feasible strategy in clinical decision-making that can ensure timely access to results and minimize the bias caused by localized tissue sampling from heterogeneous tumors. In addition, radiomics can potentially be applied to daily clinical practice to monitor responses and side reactions to ICIs.
These advantages provide convenience for clinical diagnosis and treatment, such as noninvasive biopsy, differentiating pseudoprogression from true progression, risk stratification for hyperprogression, immune-related response assessment, and so on. Current studies have indicated that the potential of radiomics in immunotherapy is substantial.
While the results of recent radiomics research are promising, they remain insufficient for its widespread application in daily clinical practice, and radiomics cannot replace biopsy or iRECIST in clinical application at this stage. Limitations and challenges in terms of practical application are not neglectable, and reproducibility presents the most significant challenge.
The heterogeneity of articles exploring radiomics features is an important issue that limits the generalization of the role of radiomics in daily clinical practice, as different imaging modalities are studied (CT, MRI, PET/CT) in different clinical settings (several different neoplasms at various stages of disease) with different "a priori" expected responsivities to immunotherapy. Due to the complexity of radiomics, few studies can be wholly reproduced, thus inhibiting the widespread use of radiomics in clinical practice (104).
Reproducibility remains a huge obstacle in the pace of clinical application Guidelines were established to standardize protocol and analysis of radiomics research (70,105). The radiomics quality score (RQS) (106) and Individual Prognosis or Diagnosis (TRIPOD) (107) were developed to bridge this gap. Moreover, ongoing single and multicenter prospective randomized clinical trials are needed to improve and validate reliability and reproducibility (Table 4). Integration and analysis of radiomics features with genomics, proteomics, and other omics data would provide additional information in precision medicine by uncovering microlevel features (105, 108).
Radiomics provides a window of opportunity for precision medicine of immunotherapy by analyzing microscopic medical imaging in a noninvasive, efficient, economical, and rapid fashion.
In the foreseeable future, we envision that radiomics will be widely applied to clinical decision-making and will serve as the impetus for the next major breakthroughs in precision medicine. However, at this stage, there are still significant challenges in the process of clinical translation and application, and further refinements are warranted.

Author contributions
HZ, QL, WW, NL, CY and LZ: Literature search, concepts development, manuscript drafts. All authors contributed to the article and approved the submitted version. All figures created with BioRender.com.

Conflict of interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.