The introduction of immune checkpoint inhibitors (ICIs) to the armamentarium of cancer therapies has led to a paradigm shift in the management of a vast range of malignancies, including non-small cell lung cancer (NSCLC). Immunotherapy uses drugs to support the immune system to combat cancer, whereas chemotherapy uses drugs to directly kill cancerous cells.
Previously, oncologists and scientists had classified these lung cancer patients into two wider categories—those who would benefit from immunotherapy, and those who would perhaps not. At present, only around 20 percent of all cancer patients can actually benefit from immunotherapy. However, with the approval of ICIs for clinical use in these malignancies, a third category of patients has started to emerge—hyperprogressors (HPs), who would be harmed by the same immunotherapy.
Hyperprogression is an atypical response pattern to ICIs described within NSCLC. The incidence varies from 4.0 percent to 29 percent, depending on tumor histology and criteria used to identify HPs (Int J Mol Sci 2019; doi: 10.3390/ijms20112674). In this subset of patients, immunotherapy paradoxically accelerates tumor growth and significantly shortens survival. The biological and clinical factors that contribute to the development of HP disease with ICIs are yet to be understood, and currently, there are no clinically validated biomarkers to identify patients at risk of hyperprogression.
New research, published in the journal Cancer Immunology Research, indicates that analysis of radiomic features on pre-treatment routinely performed computerized tomography (CT) scans could provide a non-invasive means to identify the HPs in lung cancer patients undergoing immunotherapy (J Immunother Cancer 2020; doi: 10.1136/jitc-2020-001343).
CT images of tumors contain a massive volume of valuable information in the form of subtle variations. Radiomics comprises the use of computer vision and machine learning approaches to quantitatively interrogate the subtle subvisual characteristics of radiographical images in a high-throughput method to answer pertinent clinical questions. Radiomic features have shown prognostic and predicting response to numerous different treatments across a wide range of cancers.
Studying Radiomics & NSCLC
In the current research, scientists at the Center for Computational Imaging and Personalized Diagnostics (CCIPD) at Case Western Reserve University sought to evaluate the ability of radiomic features pertaining to intratumoral and peritumoral textural patterns and tortuosity of tumor-associated vasculature to predict HP disease in patients with NSCLC treated with ICI using only pretreatment CT scans.
The team, led by Anant Madabhushi, PhD, hypothesize that HP disease would have a unique radiomics pattern associated with it when compared with other response patterns, such as responders and non-responders as determined by RECIST V.1.
Using a retrospective approach, this latest study was performed with data gathered from a total of 109 patients with advanced NSCLC who underwent monotherapy with either programmed cell death protein 1 (PD-1) or programmed death ligand-1 (PD-L1) checkpoint inhibitor drugs between January 1, 2015, and April 30, 2018, included in the study.
After implementing RESIST V.1.1 criteria, the patients were separated into responders (n=50) and non-responders (n=59). Tumor growth kinetics were used to further classify HP (n=19) among non-responders. The researchers trained computers to track and recognize patterns in preliminary CT scans taken when lung cancer is first diagnosed to reveal information that could have been useful if known before the commencement of treatment regimens. Patients were randomized into the training set (D1 =30) and a test set (D2 =79), with HPs evenly distributed between the two sets. A total of 198 radiomic textural patterns from within and around the target nodules and features relating to tortuosity of the nodule associated vasculature were extracted from the pretreatment CT scans.
The results revealed that the random forest classifier using the top features associated with hyperprogression was able to distinguish between HP and other radiographical response patterns from responders and non-responders with an area under receiver operating curve of 0.85±0.06 in the training set (D1=30) and 0.96 in the validation set (D2=79). These features included two nodule vessel-related tortuosity features and one peritumoral texture feature from 5 to 10 mm outside the tumor. Kaplan-Meier survival curves demonstrated a clear stratification between classifier predicted HPs versus non-HPs for overall survival (D2: HR=2.66, 95% CI 1.27 to 5.55; P=0.009).
Although the study is limited in the total number of HP cases in the analysis, the authors concluded that their findings “suggest that radiomic analysis of pretherapy CT scans of patients with NSCLC treated with PD-1/PD-L1 inhibitors could be used to identify patients who are at a higher risk of hyperprogression with this treatment.
Anant Madabhushi, PhD, FAIMBE, FIEEE, FNAI, Donnell Institute Professor of Biomedical Engineering; Director of the Center for Computational Imaging and Personalized Diagnostics (CCIPD) at Case Western University; and senior author on the study, spoke to Oncology Times to provide further insights into the novel study.
Oncology Times: What was one of the most significant advances in the development of this new radiomic model?
Madabhushi: “We are beginning to recognize that patients treated with immunotherapy can present with different response types. For instance, patients can be responders, non-responders, or hyperprogressors (HPs). HP is a relatively new phenomenon that is being recognized in patients treated with immunotherapy where some patients after initiation of immune-oncology (IO) go downhill very quickly.
“While there has been interest in the use of radiomics for predicting response to therapies including IO, this is one of the first studies to try to understand whether radiomics could distinguish between the different types of response—i.e., responders, non-responders, and HP. One of the very interesting aspects of the study was the fact that we invoked new classes of features, looking at textural patterns both inside and outside the dominant lung nodule on CT scans, but also looking at the tortuosity of nodule associated vasculature to distinguish the three responder classes.”
Oncology Times: Has the program been tested across different immunotherapy agents to determine if the model can predict risk of HP disease in patients with advanced NSCLC consistently?
Madabhushi: “Excellent question. While we did not explicitly look at predicting response in the context of multiple different IO agents for this specific study, we have previously looked at the utility of radiomics to predict response to IO across sites and across different drugs. In a paper we published in Cancer Immunology Research in 2020, we actually showed that our radiomic patterns distinguished responders and non-responders also predicted overall survival from CT scans in lung cancer patients across three sites and treated with pembrolizumab, nivolumab, and atezolizumab (2020; doi: 10.1158/2326-6066.CIR-19-0476). Clearly, this study needs to be repeated in the context of the HPs.
“Our previous data, however, gives us some degree of confidence that our radiomic signature will help identify HPs in a way that is agnostic to the type of agent, since these features are capturing biological attributes of the tumor and are not drug-specific features.”
Oncology Times: Immunotherapy remains extremely expensive. How will this tool help in reducing the financial stress that comes along with cancer?
Madabhushi: “Excellent point again. One of the main reasons we are working on the use of radiomics to identify the different response categories in lung cancer patients is precisely because we want to do a better job in early identification of patients who may not respond and hence might be treated by conventional therapies.
“Currently, we know that only 1 in 5 lung cancer patients tend to respond to IO, and the $200-250K price tag per patient per year, means that for every $1M spent on treating patients with IO, nearly $800K is literally going down the drain in not positively impacting the patient. One of the goals of the study is to help reduce this financial toxicity, from both a patient-centric as well as from a health economics perspective. However, beyond cost, we really owe our patients the responsibility of making sure that they are being matched with the treatment that provides the highest likelihood of a successful outcome.”
Oncology Times: What limitations of the current study remain to be addressed for this non-invasive imaging approach to be integrated into clinical settings?
Madabhushi: “Our studies thus far have been retrospective and, while they do help establish the validity of the approaches, clearly the proof of the pudding is in the prospective validation. We have started working with our medical partners to begin to design prospective clinical trials for validating the radiomic tools in a clinical setting. We are envisioning an initial limited Phase I trial, but then hope to expand out the study to a larger, multi-institutional trial. Additionally, several other related problems that we are also looking to explore, for example, distinguishing pseudo-progression from true progression and also being able to predict which patients will likely be impacted by IO-related toxicity and complications like pneumonitis.”
Dibash Kumar Das is a contributing writer.