Developing computationally derived imaging biomarkers for maximum clinical impact
Posted 16th March 2020 by Liv Sewell
Professor Anant Madabhushi is a world-recognised, award-winning leader in computerized imaging research and translational applications, with over 160 peer-reviewed journal publications and close to 100 patents issued or pending. He is a keynote speaker at the 6th Digital Pathology & AI Congress: USA. He explains here why having patents is not enough…
At Case Western Reserve University, I direct the Centre for Computational Imaging and Personalized Diagnostics. It’s a group of over 60 scientists, postdocs, faculties, students, scientific developers who are working on new AI and computational imaging algorithms for interrogating image data. Our goal is to diagnose disease presence and find subtle patterns and visual cues in routine radiographic scans and digital pathology slide images that indicate disease aggressiveness and progression, and therapeutic benefit and response.
From lab to clinic: the translational relevance of digital pathology
As a computational imaging scientist and biomedical engineer, I became interested in developing AI algorithms to develop decision support systems to help pathologists about 15 years ago. My group developed image segmentation, detection, and feature analysis algorithms to look at the shape and structure, and arrangement of cells and nuclei in pathology images. The aim was to build tools to support pathologists in diagnostic decisions and also prognosticate outcomes and therapeutic response. Some of the early projects we worked on used AI, and developed AI algorithms, to provide tools to assist pathologists in doing better screening and grading of prostate and breast cancer.
When I started to interact with clinical oncologists, I realized that the real opportunity was in leveraging these algorithms and approaches to provide decisions not so much for the pathologist, but for the clinician, and for oncologists particularly. There’s a crying need for better tools to support clinicians to identify which patients can avoid aggressive therapy, and patient response to certain therapies. These are extremely sensitive and critical decisions and unfortunately, there is a severe lack of such tools for the clinical community.
Financial toxicity is a real problem in the United States. Many cancer therapies are extremely expensive, and you cannot easily predict if the therapy will be successful. For example, immunotherapy can cost roughly $250,000 per patient, per year. But only 20-25 percent of patients receiving this therapy actually respond to it. With this situation we can see why 42% of newly diagnosed cancer patients will lose their life savings, in the US, sometimes within just 1 year of an initial cancer diagnosis.
And so, about 10 years ago, I pivoted my research direction slightly to focus on using AI and computational imaging to unearth subtle cues, patterns, and features from pathology images, alongside radiographic images, to help the clinician identify in advance which treatments the patient will benefit from and form the best treatment plan.
Leveraging AI for imaging biomarker discovery
Today, my group is developing new computationally derived imaging biomarkers. While we are using traditional deep learning or convolutional neural networks, these tools are being used to identify individual tissue and image based structures or primitives (e.g. cancer nuclei) and then using more traditional computer vision tools to extract interpretable image features. We are particularly focused on identifying computationally derived imaging biomarkers of outcome and treatment response from pathology slides that also resonate with our understanding of the pathobiology of the disease. We have developed advanced computer visual image analysis segmentation tools to identify image features with a connection to the pathobiology of the disease (e.g. spatial arrangement, shape and orientation of lymphocytes, nuclei, mitotic figures).
Last year, our group published a paper in Clinical Cancer Research showing that the spatial interplay of cancer cells and tumour-infiltrating lymphocytes was strongly associated with the likelihood of disease recurrence in early stage lung cancer. The beauty of this computational tool is that it identifies a biomarker that is too subtle and complex to be captured from the slide alone but can indicate outcome and therapeutic response from a feature that pathologists can relate to.
Another example would be some of the work we’re doing in oropharyngeal cancers where we’ve trained our algorithms to find multinucleation which is present in some tumours and indicates poor prognosis. We’ve been able to train the tool to find multinucleation instances in the pathology images and have shown that the density of these multinucleation instances is strongly associated with worse outcomes.
Demonstrating validity is a challenge, but it’s essential
One of the things that we have found to be a challenge in developing prognostic and predictive computational tools is getting access to high-quality, multi-institutional, multisite datasets with the associated long-term clinical outcome, with which to train and validate the algorithms. It’s a long process. But I’m really insistent that we have to demonstrate the validity of these approaches in a truly multi-institutional setting.
For example, when developing our algorithm for predicting prostate cancer prognosis we leveraged data from six different sites, including international sites, to demonstrate that the tool can predict patient outcome, and this took two years. But we eventually had about 1000 patient datasets which gave the tool the demonstrable robustness and validity required.
Identifying and solving problems with maximum clinical impact
As a biomedical engineer, part of my DNA is focused on seeking out questions that have translational relevance. Being a non-clinician myself I rely on my clinical colleagues to identify the unmet needs and then formulate the most impactful questions in collaboration with them. To do this I’ve built rapport with my clinical colleagues and developed the ability to have conversations with them in which I really listen to understand their perspective and the challenges they face. From these conversations we then look at how computational tools could solve the different problems and improve patient care, and how significantly. And we then choose the projects where we can have the maximum clinical impact.
An example of this process is the work we are currently doing in the immunotherapy space. Three or four years ago, the issue of patient response to immunotherapy was coming up so often in conversations with my oncology colleagues. It was apparent that there was a real need to identify which patients would respond to immunotherapy, and that a successful solution would have a huge impact. So, for the last few years a large part of our focus has been in developing computationally derived imaging biomarkers to help in predicting patient response to immunotherapy.
Clinical translation is the way I was trained, it’s the way I think. As an immigrant from India, I’m driven to translate across settings too. For example, in India, many patients with early stage breast cancer can avoid chemotherapy but the screening test is too expensive for most. I’m involved in developing a low-cost test with the Tata Cancer Centre in Mumbai. Developing low-cost computational pathology solutions for pain points in resource-scarce settings is another example of the maximum clinical impact projects I’m working on.
The goal: better patient care
Ultimately, the dream, and my goal, is for discoveries to enable better patient care in the clinic. You can get the paper out, even have patents, but it doesn’t mean anything unless it’s been deployed for the benefit of patients. Providing solutions for low- and middle-income countries is also particularly gratifying for me personally.
Anant Madabhushi is the F. Alex Nason II Professor of Biomedical Engineering at Case Western Reserve University, Cleveland, Ohio and Research Health Scientist at the Cleveland Louis Stokes Veterans Affairs Medical Centre.
The 6th Digital Pathology & AI Congress: USA will provide a collaborative environment to advance the adoption of digital pathology, explore the latest clinical applications, and engage with case studies. Find out more.
Leave a Reply