Written by Audrey Woods
Image analysis is at the core of almost every branch of medicine. Neurologists read brain scans to diagnose strokes, oncologists use MRIs to investigate soft tissue maladies like cancer, and general practitioners look at X-rays to confirm broken bones. In time-sensitive situations such as heart interventions or fetal imaging, accurate, rapid, and precise analysis of medical images can be critical to saving lives.
Nowadays, advancements in medical imaging, especially when combined with AI and other computer science methods, are opening new frontiers in various specialties. For instance, radiologists equipped with AI models can analyze images faster, and surgeons can use 3D imaging and computer-aided design to visualize complex anatomies and plan more precise interventions. In short, computer science applied to medical image analysis enhances the capabilities of healthcare professionals, broadens access to high-quality medical care, and improves outcomes across the board.
For this reason, CSAIL Professor Polina Golland has dedicated her career to developing novel techniques for biomedical image analysis, pursuing solutions to real-world problems and advancing medical science in the process.
FINDING HER INTEREST
Professor Golland was drawn to computer science early in her career because, as she explains, “I always liked mathematics, and computer science offered a way to do mathematical analysis in the context of practical problems.” When she started graduate school at MIT, she entered with a focus on computer vision. But when she saw some of her peers working on medical images, she was curious. Working in such a targeted way on problems in a real-world domain excited and fascinated her. “There is also a natural human interest in medicine,” she says, an interest that also attracts many of the researchers in her lab. Professor Golland says, “once I tried working on medical images, I never looked back.”
Since joining the MIT faculty in 2003, Professor Golland and her Medical Vision Group have been focused on developing algorithms for medical image analysis and visualizations of medical imagery, collaborating extensively with clinicians, clinical researchers, and neuroscientists to make sure these methods can be applicable in the field. For her outstanding contributions in this area, she was named a fellow of the American Institute for Medical and Biological Engineering (AIMBE) in 2021. With AI and ML rapidly gaining traction in healthcare and biology, the practical applications of Professor Golland’s research show no sign of slowing down.
THE MEDICAL VISION GROUP: AI MODELS FOR HEALTHCARE IMAGING
One of the areas Professor Golland and her group work on is the challenge of fetal imaging, particularly with MRI. MRI works by taking a series of cross-sectional images and splicing them together to make a 3D reconstruction. It’s important when taking an MRI that the subject doesn’t move between slices, or the 3D reconstruction won’t be accurate. This is easy with most adults, who understand how to sit still for the 15 minutes or so it takes to do an MRI scan, but problematic when trying to take an image of a moving fetus. Fetal MRIs require more than double the time, as the technician tries to “chase” the baby around and capture accurate images, and also can be slow to analyze, since radiologists have to painstakingly compile potentially mismatched images. In situations where decisions must be made quickly about a developing fetus, such a long process is not ideal.
To address this problem, Professor Golland’s group is working on two different but complimentary methods. First, they’re developing methods such as the one presented in a 2024 paper which use AI to reconstruct the images significantly faster. Their model uses a fully convolutional network to predict how the slices of a target fit together, requiring only one stack of images and less time in the scanner. Another project in this area, even more exciting to Professor Golland, is their current work developing an algorithm that would identify the head inside the scanner and then, in real-time, do the “chasing” of the fetus for the technician, re-aligning with the head and determining where the next slice should be. This method could also be useful in, for example, taking MRI scans of elderly patients with Alzheimer’s or Parkinson’s disease who have trouble staying still for the duration of the scan. The best part is that the software they’re developing is complimentary with the current equipment, meaning it can easily be added to existing scanners.
Another area of Professor Golland’s research is enhancing medical imaging for intervention support. Currently, her group is working to aid interventions that use imaging to navigate in real time. Take, for example, intravascular procedures, where physicians generate detailed 3D scans of a network of blood vessels in a given area to identify the issue. Then, to perform the intervention, the patient is injected with a contrast agent that shows up in an X-ray which a physician uses to guide the catheter to where treatment is needed. The issue is that X-ray only provides 2D images, so the doctor has to navigate in 3D space with only 2D guidance, which can lead to errors and waste valuable time.
“What we are working on,” Professor Golland says, “is taking the 3D imaging from before that has all the information and integrating it with the 2D information that we have during the procedure.” A recently published paper tackles the first hurdle in this challenge, presenting a method which finds spatial correspondences between the 2D image and the 3D image, linking the two together such that a physician might be able to click on a part of the 2D image and it would be highlighted on the 3D image, or vice versa. Going forward, this project aims to, as Professor Golland explains, “lift the 2D information and enhance the pre-operative 3D information with what we're seeing in real time in 2D.”
The third broad category of Professor Golland’s research is using AI for medical image labeling and analysis. She says this research started because of a need to quickly read, for example, chest X-rays. In a typical emergency room setting, there isn’t time for physicians to go through all the images that an X-ray might produce. But if the images could be numbered, like blood glucose or hemoglobin, that could offer a shorthand to identify many common issues. This would also be useful in clinical trials, which require quantified results but are currently burdened by the cumbersome task of physicians needing to do the quantifying.
However, this research has since evolved to address the need for AI that can understand and annotate medical imagery. Her group is developing algorithms that can “read” x-rays, verbally describing a given area or pointing to the area in question when prompted by text. Such methods—which combine language models and image models—can not only label medical images, they can also suggest the presence or absence of things in the image, and even have the potential to predict larger medical questions, like whether a patient is likely to return to the hospital shortly after being discharged.
While Professor Golland is excited to be solving real-world problems, she acknowledges that it takes a long time for her work to be implemented into clinical practice. She explains, “people who study cryptography or invent new encryption mechanisms don’t put the algorithm into the ATM.” Because of that, Professor Golland enjoys working with the startups and companies who come to her for guidance on the commercialization aspect, particularly if they have an application that will improve patient outcomes. Ultimately, the thrill of her research stems from the critical constraints of medicine. Working in a field where a correct answer or functional tool could literally save lives makes Professor Golland excited to continue what she’s doing.
LOOKING FORWARD: FUTURE RESEARCH & ADVICE FOR INDUSTRY
Collaboration is a necessary component of Professor Golland’s research, which cuts across many domains. She and her group work with doctors, clinical research institutions, corporations, and startups to make sure their efforts are being put toward practical questions. She elaborates, “if somebody sends me an email and says, ‘I have this challenging problem and it has to do with medical images, can we meet and talk,’ I say, ‘sure,’ because I love learning about new applications of medical imaging.”
One common theme she’s seen in her role as an advisor, both to students and startups, is that complex and sophisticated computer architectures rarely outperform simple architectures trained with lots of data. There are many creative and innovative models out there, but Professor Golland has found that currently, nothing can compete with injecting more data when it comes to improving accuracy. This is why using machine learning to annotate or label medical images—which is the most time-consuming and expensive part of healthcare research—is so exciting.
When asked what she wants readers to know, Professor Golland answers, “how awesome my graduate students, postdocs, and collaborators are.” A huge reason she enjoys her work so much is the “amazing group of people I'm working with.” Their enthusiasm, combined with her own, creates the perfect environment to push the boundaries of biomedical image analysis, empowering better, more efficient, and life-saving medical practices.
Learn more about Professor Golland on her website or CSAIL page.