NEW YORK – Researchers in the US and China have developed an imaging-based deep learning model for predicting EGFR mutation status and guiding treatment decisions for non-small cell lung cancer patients — a method they describe as a non-invasive alternative to tissue-based assays.
For NSCLC patients, a positive EGFR mutation test is key to determining their eligibility for treatment with tyrosine kinase inhibitors, or, if they test negative, for immune checkpoint inhibitors. But currently marketed tissue-based PCR and next-generation sequencing tests to gauge EGFR status can be a hindrance since advanced cancer patients commonly do not have enough biopsy tissue available to facilitate this assessment.
Moreover, tumors are heterogeneous, and an EGFR mutation detected in a tissue sample may not necessarily reflect a patient's tumor characteristics. EGFR mutation status can change, too, as patients' cancers progress and their tumors become resistant to various treatments, but repeat testing to capture these changes may not be an option for patients with insufficient tumor tissue.
Blood-based molecular diagnostics that gauge EGFR mutation status avoid the pitfalls of tissue testing. However, turnaround times for both types of tests can be days or weeks, resulting in treatment delays for the patient.
After considering these limitations, the researchers set out to develop an imaging-based, radiomics strategy and ultimately created a deep-learning model, published in Nature Communications last month, they believe to be a faster and less invasive option than current biomarker testing methods.
According to co-senior author Matthew Schabath, a researcher at the H. Lee Moffitt Cancer Center and Research Institute, the algorithm developed by his team is able to classify an image from PET or CT scans and provide an EGFR score in real time. "The advantage of doing these sorts of analyses on images is that as soon as the image is acquired, a computer could hypothetically start calculating whatever the score of interest is," he said. "In this case it's an EGFR deep learning score."
While the use of imaging to guide treatment decisions is not new in oncology, Schabath said that the types of analyses detailed in the paper have only become possible in the last decade due to the rise of radiomics, which relies on computer algorithms to analyze high-dimensional imaging data.
"The term radiomics has only been around for seven or eight years," he said. "While this isn’t the absolute first study to look at imaging correlated with EGFR, this is a relatively new research area. We’re one of many groups on the forefront of this whole field of [applying] neural networks, deep learning, machine learning, and computer science to medical research."
The researchers' radiomics-based method, which they developed and validated retrospectively on multiple cohorts of patients from China and the US, uses a computer-generated, deep-learning score, dubbed EGFR-DLS, designed to predict patients' responses to treatment with TKIs or immune checkpoint inhibitors, respectively. The score is generated based on analysis of PET/CT images, which are run through the algorithm.
Prior to the imaging, patients receive a fluorodeoxyglucose (FDG) tracer, the uptake of which, according to the researchers, is "known to be affected by EGFR activation and inflammation." This means that on a PET scan of an NSCLC patient who has received the FDG tracer, EGFR-mutated tumors will be distinguishable from EGFR-wildtype tumors. Because FDG-based PET imaging is commonly used to stage lung cancer, the researchers were able to access cohorts of patients who had undergone this imaging and retrospectively assess their treatment outcomes, as well as the consistency between the EGFR-DLS and their known EGFR mutation status assessed by tissue-based testing.
Algorithm training, validation
To develop the EGFR-DLS, the researchers retrospectively analyzed more than 600 patients treated at two hospitals in China. They validated the score retrospectively on a third cohort of Chinese patients. Finally, to assess the score's ability to guide treatment decisions, the researchers relied on a cohort of NSCLC patients who had received TKIs at the Fourth Hospital of Harbin Medical University in China and a cohort of patients who had received immune checkpoint inhibitors at the Moffitt Cancer Center. In total, PET/CT imaging data from 837 NSCLC patients were used to develop and validate the EGFR-DLS.
In the training, validation, and external test cohorts, the researchers reported the accuracy of the EGFR-DLS score as 81 percent, 83 percent, and 79 percent, respectively.
When researchers used the algorithm to stratify NSCLC patients into those with high and low EGFR-DLS scores, they found that patients treated with TKIs and high scores tended to have longer progression-free survival compared to those with low scores.
Moreover, higher EGFR-DLS scores also appeared to be associated with better TKI response. Among 31 patients who saw their tumors shrink following TKI treatment, the median EGFR-DLS score was 0.53. On the other hand, among 36 patients who experienced stable disease or disease progression following TKI treatment, the median EGFR-DLS score was 0.38.
When it came to the relationship between patients' EGFR-DLS scores and their responses to immune checkpoint inhibitors, the researchers found the opposite to be true. Immunotherapy-treated patients with low EGFR-DLS scores had a median progression-free survival of a year, compared to four months among those with high scores.
Additionally, 67 percent of immunotherapy-treated patients with low scores experienced a durable clinical benefit (defined as progression-free survival for more than six months) versus 33 percent of patients with high scores. Furthermore, 33 percent of immunotherapy-treated patients with high scores experienced hyperprogression (defined as time-to-treatment failure of less than two months) versus 16 percent of patients with low scores.
Alongside the EGFR-DLS, researchers also developed a PD-L1 deep learning score, dubbed PDL1_DLS, which they determined was similar to immunohistochemistry-based PD-L1 testing in terms of prognostic value. Combining the EGFR-DLS and PDL1_DLS, and assessing both signatures alongside patients' outcomes, they found that those with low EGFR-DLS and high PDL1_DLS scores experienced significantly longer progression-free survival with immune checkpoint inhibitors versus patients with these scores who were treated with TKIs.
The outcomes further suggested that patients with both low EGFR-DLS and low PDL1_DLS may be poor responders to both TKIs and immune checkpoint inhibitors. For these patients, Schabath noted, standard chemotherapy might be the best option.
A limitation of the PD-L1_DLS score, according to the researchers, was that PD-L1 status was known for only 75 percent of patients in the immune checkpoint inhibitor-treated cohort. As such, "the complementary information of EGFR-DLS in guiding immunotherapy needs to be validated on a larger cohort with PD-L1 status," they wrote.
Ultimately, based on the relationships the researchers observed between the deep learning scores and clinical outcomes, the team was able to demonstrate that the EGFR-DLS is both predictive and prognostic. "Many retrospective trials are really only prognostic," Schabath said. "But we also show it's predictive because it can help with the decision between using a TKI or immunotherapy or standard chemotherapy."
Next steps, potential implementation
The retrospective nature of the latest study necessitates further prospective validation before oncologists can consider adopting this method for making treatment decisions. In future studies, the researchers hope to assess the applicability of this method longitudinally over the course of patients' treatment and validate it in larger prospective trials, Schabath said.
"Our next steps will be determining how we can get this into a clinical trial," he said. "Perhaps we'll start with an observational trial and, depending on our metrics, we can move into a randomized trial to validate this work."
Schabath's group hasn't ironed out the design details of a potential randomized study design yet, though. The question they want to answer — whether the algorithm predicts treatment benefit or accurately determines biomarker status — will influence the design.
"Our biomarker is a very good surrogate for EGFR [mutation status]," he said. "If I'm putting on my epidemiology hat, I would say we would want to randomize according to treatment … But this is something that we'd have to [discuss] with medical oncologists and clinical trialists to figure out the best study design."
Noting the early nature of the present research and the need for prospective validation, Schabath outlined how this radiomics-based method might be adopted clinically.
Although the researchers required specialized technology with a great deal of computing power to develop the deep-learning score algorithm, Schabath said that the algorithm itself could be programmed into any radiology lab that has PET and CT equipment. The specific process of programming the algorithm into a picture archiving and communication (PAC) system would need to be developed and fine-tuned, but according to Schabath, existing software can facilitate this integration.
"This is not something that is 'plug and play' right now, but it wouldn't be very difficult to be able to create the infrastructure to be able to do this," he said. "Any radiology center would theoretically be able to have this [algorithm]. You would put the algorithm into the back end of the PAC system, and it would provide scores back to you about risk or treatment or prognosis."
Despite the feasibility of this approach, there are several potential barriers to the broad uptake of this type of technology. Radiology centers with PET imaging technology are not widely available to cancer patients in all areas of the world. In addition, clinical imaging is expensive for patients, their providers, and health systems. The study authors acknowledged that "this model may be limited to the developed countries and to large urban centers in the developing countries."
Schabath recognized these hurdles in radiology and highlighted efforts to develop more accessible, bedside, or handheld PET imaging technologies. He further noted that although the specific algorithm detailed in this study was developed on PET, the signatures could be based on standard CT imaging, too. In contrast to some biomarker assays that saddle patients with out-of-pocket costs, Schabath also suggested that the path to reimbursement for this type of algorithm may be eased by the fact that PET imaging is already part of standard-of-care in many settings.
One advantage of the imaging method is its noninvasiveness relative to tissue-based assays. Indeed, the paper's title, "Noninvasive decision support for NSCLC treatment using PET/CT radiomics," places the spotlight on this point.
Schabath slightly amended this terminology when talking about it, however. "It's minimally invasive, to be appropriate about it," he said, explaining that, even in the absence of a required biopsy, imaging is not entirely free of invasive elements. "When we claim imaging is not invasive, I often hear from patients [about] getting exposed to radiation," he said. "And when we talk about PET, you're getting a PET tracer."
It also takes time to go into a radiology lab for PET imaging, and if this type of test were to be repeated longitudinally throughout the course of a patient's treatment, further research would need to establish an ethical framework to make sure that patients are not being exposed to excess radiation, Schabath acknowledged.
Still, he is optimistic about the future of radiomics for this type of assessment and its utility beyond EGFR and lung cancer. In his own institution, for instance, similar work has been done with KRAS and ALK biomarkers. "These neural networks can be trained and tested and validated with large datasets that have images and patient data for whatever mutation it is," he said. "It doesn't matter what sort of cancer it is."