Deep Learning and the Future of Pulmonary Fibrosis Research
Muhunthan Thillai, Chief Executive Officer & Co-Founder, Qureight
In this Q&A, we speak with Muhunthan Thillai, CEO and Co-founder of Qureight, a company at the forefront of applying artificial intelligence to medical imaging. Our discussion focuses on how their AI deep-learning imaging technology is changing clinical trials for progressive pulmonary fibrosis (PPF), offering new possibilities for diagnosis, treatment, and research.
It highlights some of the key advancements taking place in the space, such as 3D imaging, outcome prediction, and the creation of synthetic study arms, demonstrating a significant leap forward in the fight against PPF.
1. How is AI deep-learning imaging technology being integrated into clinical trials for progressive pulmonary fibrosis (PPF), particularly in the MIST study?
Deep learning imaging technology is being integrated into clinical trials to more deeply examine structures of the lung, and map treatments to disease outcomes. The MIST study, a global lung fibrosis study conducted by Avalyn Pharma, is collecting data from over 200 patients worldwide and studying the effects of AP01 (inhaled pirfenidone) in PPF patients. In this study, AI deep learning technology is being integrated to analyse CT scans, improve understanding of the treatment, and curate key data points for understanding the disease.
The AI imaging technology will examine lung structures such as airways and blood vessels. By comparing changes in these structures over time, and between the drug and placebo groups, researchers can assess the drug's effectiveness. This analysis will also provide insights into how the treatment affects different lung structures, potentially providing evidence to advance the trial.
2. What specific advantages does deep-learning analysis offer over traditional radiologist-based interpretation of HRCT scans in PPF patients?
Traditional biomarkers for pulmonary fibrosis (PPF) include lung function tests, where patients breathe into a machine to measure lung capacity. While these tests are useful for large-scale studies, they can be unreliable for individual patients due to variability in breathing effort, daily health, or machine setup. Another traditional method is a qualitative assessment of scans by radiologists, who evaluate changes in fibrosis. However, this approach lacks precise quantification.
Deep learning offers a more objective and quantitative alternative. By analysing CT scans, deep learning extracts 3D data and provides accurate classifications of lung structures like fibrosis, airways, and blood vessels in real time. This provides trial sponsors with clear, numerical data on the changes occurring within the lungs, offering a distinct advantage over both subjective radiologist interpretations and the variable results of lung function tests.
3. How do AI algorithms enhance diagnostic accuracy in classifying fibrotic lung diseases, and what are the key challenges in their implementation?
The key advantages of AI technology are its reproducibility and speed. While a radiologist might spend seven or eight hours meticulously outlining airway structures on a single scan and then subjectively assessing their size, deep learning models offer a significantly faster and more consistent approach. Built using 3D convolutional neural networks (3D CNNs), these models essentially mimic the collective expertise of multiple radiologists, achieving a high level of accuracy. A recent study that we published in American Journal of Respiratory and Critical Care Medicine (AJRCCM) showed that our 3D models achieved a DICE score of over 85%, indicating they are more than 85% as accurate as a group of highly trained radiologists. This combination of accuracy and speed is a crucial advantage. However, implementing this technology within clinical trials presents challenges. Simply having an algorithm isn't enough; you need to have the infrastructure to deploy it. That’s why deep learning platform companies have been focused on building the necessary infrastructure to deploy this technology successfully within clinical trials over the past few years.
4. In what ways has deep learning demonstrated improved outcome prediction in patients with progressive fibrotic lung disease, and how does this impact trial design?
The primary benefit of this technology is its ability to provide highly accurate measurements of lung conditions, far exceeding the precision of traditional methods. This advancement holds significant promise for the next generation of clinical trials. Currently, patients undergo scans at the start (baseline) and end (e.g., after six months) of a study. We can now analyse these baseline scans to quantify the disease burden and predict patient outcomes.
In our study published in AJRCCM, deep learning technology could accurately predict patient trajectories 12 to 24 months into the future. This predictive capability enables us to 'cohort enrich' trials, selecting patients at baseline who are most likely to exhibit disease progression. This is crucial because these trials require participants with a certain level of disease that progresses predictably. If participants have mild, non-progressive disease, it becomes challenging to determine the drug's efficacy. Therefore, cohort enrichment and patient stratification have significant implications for trial success, cost-effectiveness, and overall efficiency.
New imaging models, validated in a large global cohort, have shown that quantifying baseline fibrosis levels allows for accurate 12-month outcome predictions. One of our deep learning algorithms can identify patients with a baseline fibrosis level exceeding 12% as likely to progress within the year, while those below 12% are less likely to progress. These absolute fibrosis thresholds can help predict who will worsen and who is best suited for clinical studies. This technology, previously unavailable, is now being implemented in trials.
5. How does the 3D based methodology applied in deep-learning models improve the objectivity of fibrosis assessment compared to conventional 2D-image-based analysis?
Until about four or five years ago, we primarily relied on 2D approaches for analysis. These 2D methods examine square pixels in lung scans to identify abnormalities. This method was based on support vector machine learning, a technology that was widely used in research, not just for lung disease but for many other conditions as well.
However, advancements in brain imaging research led to the development of 3D convolutional neural networks. Instead of 2D slices, these networks construct 3D models of the brain. This 3D approach has now become standard in brain scan analysis and deep learning. We adapted this 3D technology to lung imaging, leveraging insights from our colleagues and partners in the brain imaging field, and the 3D models offer significantly improved accuracy. In recent studies, researchers compared the 2D and 3D technology on the same scans, and the results clearly showed greater accuracy in identifying lung structures with the 3D method.
6. What role does AI-driven imaging play in accelerating drug development for PPF, and how might it influence regulatory approval processes?
AI-driven imaging can accelerate drug development for PPF in two key ways. First, it enables faster and smaller trials. Recent work with our pharma partners shows that by using scans instead of traditional breathing tests, the study results could have been replicated with 40% of the original number of patients, reducing the time and resources. Additionally, a six-month study could potentially be reduced to three months using imaging technology. This demonstrates that deep learning imaging technology allows for both smaller and faster trials, reducing costs and getting treatments to patients quicker.
We're starting to show that deep learning imaging technology can not only allow for smaller trials but also faster trials. In terms of regulatory approval for treatments developed in this way, we're still in the early stages. There are currently no FDA-approved biomarkers for lung imaging, but we anticipate the first approvals within two or three years, and we're confident this use of technology will be among the first. To reach that point, we need to conduct numerous large-scale clinical trials over the next few years.
Second, selecting patients for lung fibrosis trials can be challenging due to the prevalence of standard-of-care treatments. Finding patients who are treatment-naive, meaning not currently on these standard treatments, is difficult. Most patients in these studies are already receiving standard drugs, which complicates data analysis. While some patients are not yet on medication, they are fewer in number. Therefore, the ability to conduct smaller trials and specifically select treatment-naive patients would be highly advantageous for pharmaceutical companies.
7. What future advancements in AI deep-learning imaging technology are anticipated, and how might they further transform clinical trials and patient care in progressive pulmonary fibrosis?
As well as reducing time and costs of clinical trials, there’s an exciting future ahead where AI can create synthetic or ‘digital twins’ of clinical studies, allowing even more in-depth comparison between patient groups and medical outcomes. For example, AI can create synthetic matched placebo groups, where we can compare treatment with a simulated outcome. Although it’s early for this kind of work, we have already done this in practice with some of our partners. Since deep learning platforms curate a vast amount of longitudinal data across a study, we can utilise this to build a granular picture of disease progression under different parameters.
In lung fibrosis trials, we use curated data from the drug arm of a study, such as patients' baseline CT scans, comparing initial levels of fibrosis (fibrosis quantification) to identify patients on our platform with baseline CT scans similar to those in the study group. We then used AI models to create a matched placebo group, indicating the likely outcome or progression of patients without the treatment.
By using baseline CT scans and fibrosis markers to carefully match patients, we were able to create a synthetic matched placebo arm, demonstrating the likely outcome had they used an imaging-matched control. This level of granularity is rare in clinical trials and offers far greater accuracy than simply selecting placebo patients. It's a genuinely exciting time for the sector, as these advancements represent a quantum leap forward in our fight against life-threatening lung conditions like progressive pulmonary fibrosis.
Author Bio
Dr. Muhunthan Thillai is a highly experienced pulmonologist with over 20 years of expertise in thoracic medicine. He graduated from Imperial College London, trained at Oxford, and is a member of the Royal College of Physicians. He holds a PhD in molecular immunology and was previously the Director of Interstitial Lung Diseases at Cambridge hospitals. He is currently an Honorary Associate Professor at the University of East Anglia.
