[¹⁸F]FDG PET immunotherapy radiomics signature (iRADIOMICS) predicts response of non-small-cell lung cancer patients treated with pembrolizumab

In spite of the advances in lung cancer treatment, prognosis for patients has been poor with a 5-year survival rate around 15%.1 A new hope has come with renaissance of immunotherapy, such as programmed death-1 antibodies (anti-PD-1), which invigorate a patient’s immune system to fight against malignant cells.2 In non-small-cell lung cancer (NSCLC), which represents 85% of all lung cancer cases, treatment outcomes of anti-PD-1 immunotherapy are significantly better compared to conventional cytotoxic therapies. In selected patient population, response rates can be over 40%.3 The responding patients usually achieve durable benefit and prolonged survival. Occasionally, even complete remissions of metastatic disease are observed, but such complete responses are still in minority.

Due to possible unusual response patterns (e.g. pseudoprogression), treatment response assessment in immunotherapy is challenging.4 The most routinely used methods are Response Evaluation Criteria in Solid Tumours (RECIST) and its modification for use in immunotherapy (iRECIST), among others.5 Although iRECIST was found superior to RECIST in identifying pseudoprogression, iRECIST is a late response assessment method, because anatomical changes observed on computed tomography are usually delayed, and the suspicion of progressive disease needs to be confirmed with an additional scan 1–2 months after the first assessment.6 Importantly, studies have shown that none of the RECIST-based endpoints could be used as valid surrogates for overall survival (OS) in anti-PD-1 trials, while the correlation of iRECIST-based endpoints with OS is yet to be explored.7, 8 Since the molecular and functional tumour changes are known to appear faster compared to anatomical changes, several immunotherapy response assessment methods, based on 2-deoxy-2-[fluorine-18] fluoro-D-glucose positron emission tomography/ computed tomography ([18F]FDG PET/CT), have been proposed.9, 10, 11, 12 However, there is still a lack of sufficient evidence to infer, which method, if any, might be the most appropriate for the routine clinical use.13, 14, 15

Recently, research into the identification of new biomarkers for use in immunotherapy has also increased. Various predictive and prognostic biomarkers of response have been identified, including tumour PD-1 ligand (PD-L1) expression, tumour mutation burden, tumour infiltrating lymphocytes density, mismatch repair deficiency, microsatellite instability, and gut microbiota.16, 17 However, the reports from different studies sometimes oppose each other, therefore the current biomarkers need further validation.18 Moreover, most of them require invasive biopsies, and are impractical or too expensive for a routine clinical use. On the other hand, few immunotherapy clinical studies examined possible non-invasive imaging biomarkers, but there is still a lack of research performed in NSCLC patients.14 Three retrospective anti-PD-1 studies showed associations of pre-treatment sum of maximum standardized uptake values (SUV_max) of all lesions (SUV_maxwb)19, SUV_max of the most avid lesion20, and volumetric parameters (metabolic tumour volume [MTV], and total lesion glycolysis [TLG])21, with NSCLC patient response as defined by RECIST. However, significant correlations of these features with OS were not observed. There is also a lack of clinical studies in immunotherapy investigating more sophisticated image analysis methods such as radiomics analysis. Radiomics analysis harnesses the full power of medical imaging by extracting numerous quantitative features, hypothesized to reflect more deeply the tumour phenotype, as well as the genotype.22, 23 Recent anti-PD-(L)1 radiomics studies have shown associations of CT radiomics signatures with tumour immune phenotype24, hyperprogression25, and progression-free survival (PFS)26. Moreover, two studies also examined the predictive value of PET radiomics features. Polverari et al. observed significant differences in tumour heterogeneity (as defined by kurtosis and skewness) between patients with progressive disease (PD) and non-PD21, while the study by Mu et al. proposed a combined PET and CT radiomics signature for predicting patient PFS and OS.27 In these studies (except Polverari et al. ), data mining using vast number of features (up to 1160) was performed in order to build multivariate radiomics signatures containing up to eight features. Although on one hand, such approach might allow for a more precise quantification of tumour characteristics, on the other hand, the so obtained predictive models could be prone to overfitting, and probably too complex and non-intuitive for a successful clinical translation. Moreover, it is also well known that a lot of radiomics features are not suitable candidates for biomarkers, for example due to an excessive test-retest variability.28

The primary aim of our prospective study was to determine whether immunotherapy [18F]FDG PET radiomics signature (iRADIOMICS) predicts response of stage IV NSCLC patients to pembrolizumab better than the current routinely used clinical standards (PD-L1 immunohistochemistry, and iRECIST). To overcome the aforementioned pitfalls, we deliberately analysed only a small subset of radiomics features, which were previously proven to be robust and reliable according to test-retest variability28, and built iRADIOMICS with minimum number of features.

New biomarkers of response to immunotherapy are urgently needed. In NSCLC, PD-L1 TPS is still the only predictive biomarker routinely used in clinics, in spite of the growing evidence suggesting that it is far from optimal.36 Among the reasons for its questionable predictive power are inconsistent measurement methodologies, intratumour PD-L1 expression heterogeneity, and the fact that immune cells infiltrating the tumour can express PD-L1.37 Even in our study, the survival predictions based on PD-L1 TPS performed poorly. Additionally, because it is not clear to what extent the current standards for treatment response assessment (RECIST, iRECIST) correlate with overall survival (OS), the duration of treatment, as well as the decision about cessation of anti-PD-1 immunotherapy, rely on the subjective judgment of the treating physician, which is mainly based on the observed immune-related adverse events and achieved clinical benefit.

We aimed to address these issues with the use of [18F]FDG PET/CT imaging, since it is widely used, affordable, and non-invasive. When we examined the predictive ability of individual radiomics features, we found that some of the features showed high predictive power at baseline, while at month 1 and month 4 their informative value decreased significantly. This is consistent with a number of studies suggesting that intrinsic tumour characteristics, such as tumour histopathology, tumour microenvironment, and immune contexture, most likely have a major impact on response to immunotherapy.16, 17,38 The most dominant feature was Small Run Emphasis (SRE), which was able to discriminate responders from non-responders to anti-PD-1 therapy, it had a significant relationship with patient OS, and high predictive power. In patients with SRE > SRE_median, the probability of survival by Kaplan-Meier analysis was also significantly higher. Although studies have shown that texture-based features might reflect tumour heterogeneity on macroscopic, cellular, or even molecular or genomic level39, their clear relationship with the underlying biology still needs to be elucidated. However, from the definitions of texture features used in our study we can infer that at baseline, primary tumours of responders have finer and more homogeneous metabolic structure, as reflected by higher SRE and lower Entropy-GLCM, respectively. See Table S1 for formal mathematical definitions, as well as intuitive descriptions of the studied texture features. In terms of underlying biology we could speculate that these findings might reflect tumours with spatially more homogeneous clonal structure, more homogeneous intrinsic infiltration of immune cells, more homogeneous tumour microenvironment, or fewer hypoxic or necrotic regions. Interestingly, this finding is in agreement with the study by Polverari et al., where patients with progressive disease (PD) exhibited higher tumour heterogeneity at baseline (reflected by higher kurtosis and skewness), compared to non-PD patients. On the other hand, the finding is at odds with the study by Mu et al., where heterogeneous tumours presumably had a higher chance to achieve durable clinical benefit.27 However, heterogeneous tumour phenotype in this study was inferred from two components of eight-variable radiomics signature, making intuitive conclusions about the underlying tumour biology even more difficult compared to our study. In agreement with the study by Takada et al., we also observed the trend of higher SUV_max among the responding patients, although it was not statistically significant.20 A similar lack of statistical significance of SUV_max, or even the opposite trend, was observed by other groups, therefore the predicitve value of SUV_max should be considered highly questionable.13,19,21

We analysed only primary tumours, yet neglected lymph nodes (LN) and distant metastases (DM). The main reason for this approach is that radiomics analyses might not accurately quantify intra-tumour heterogeneity of small lesions due to the partial volume effects, which could be even more pronounced in PET imaging with limited spatial resolution.40 However, inclusion of LN and DM in future predictive models could additionally improve their predictive power and accuracy. Especially an [18F]FDG PET signal of LN might be connected with the cancer immunity cycle, possibly capturing the processes that occur in LN after the initiation of anti-PD-1 therapy, including T cell priming and activation.41

The analysis of the predictive ability of iRECIST, PD-L1, and iRADIOMICS signatures revealed some interesting aspects. First, the response to anti-PD-1 therapy seems to occur fast, as iRECIST signature was able to predict the response of 76% of patients already at month 1, while the predictive ability at month 4 had not improved. These results suggest that treatment response assessment could be performed as soon as 1 month after treatment initiation. Moreover, its satisfactory ability to predict OS indicates that clinical decisions about (dis) continuation of anti-PD-1 therapy could (at least in part) rely on iRECIST assessment rather than purely on the observed clinical benefit. However, the correlation of other iRECIST-based endpoints with patient survival should be further explored.

Lastly, the iRADIOMICS was found superior to PD-L1 and iRECIST both in terms of predictive power and, importantly, timing. From the clinical point of view, each additional month (or day) of an ineffective therapy can be crucial for metastatic NSCLC patients. The fact that the iRADIOMICS was able to correctly predict the response of almost 80% of patients before therapy, could have an important clinical impact. The predicted non-responders to pembrolizumab could be offered other treatment options to improve their OS. However, the predictive ability of iRADIOMICS needs to be confirmed in future independent studies with a higher number of patients.

Our study compared the predictive power of baseline biomarkers (iRADIOMICS and PD-L1) to the early treatment response assessment method (iRECIST) – single point vs. multiple point assessment. However, from the practical standpoint, the baseline prediction is desirable to the treatment response assessment as it is earlier and allows more time for favourable clinical decision making. Potentially the two approaches could be combined, but such study would require higher number of patients to secure clinical significance because of more degrees of freedom (variables).