Written by Matthew Busekroos
Marianne Rakic is a fourth year PhD student working in the Clinical and Applied Machine Learning Group at CSAIL.
Originally from Belgium, Rakic completed a MSc in Electrical Engineering and Information Technology at ETH Zürich prior to MIT. She received her Bachelor's degree in Electrical Engineering at the University of Liege. Rakic is passionate about computer vision for medical imaging and its different applications in healthcare.
Rakic’s most recent project, Tyche, is a medical image segmentation model that aims at generalizing to new tasks and capturing uncertainty in the medical image.
“In biomedical imaging, segmentation consists in annotating pixels from an important structure, for example an organ, a lesion, or a cell part,” Rakic said. “It is a central task for both biomedical researchers and clinicians, and is typically solved with neural networks.”
Rakic said together with their biomedical collaborators, they have identified two fundamental shortcomings with existing methods.
“First, for most new segmentation tasks (different regions of interest or different imaging types or protocols), a new model has to be trained,” she said. “This requires extensive resources and machine learning expertise, which is often infeasible for medical researchers and clinicians. Second, most existing segmentation methods produce a single segmentation for a given image,” Rakic added. “In practice however, there is often considerable uncertainty, and different expert annotators will often segment the same image differently.”
“We tackle both of these problems with Tyche, a segmentation tool that both generalizes to new tasks without the need of machine-learning expertise and generates multiple plausible segmentation candidates that capture the inherent uncertainty in an image,” she said. “Tyche can be used on a wide set of tasks out of the box and can help enable studies of human disease that were not possible before.”
Rakic said that models like Tyche can help clinicians and researchers capture the inherent ambiguity of their data and limit their dependency on model retraining.
“Basically, with Tyche, we propose a solution to those two challenges jointly. It is a single model that you train once. It adapts to different modalities and anatomies, and captures uncertainty by producing different candidate segmentation,” Rakic said.
Rakic said one last aspect she really likes about Tyche is the loss function used to train it.
“The way we teach our AI model to output multiple diverse predictions is by focusing on the best candidate at every training iteration” Rakic said. “If the model outputs four segmentations at training time but only one of them is good, we teach the model to focus on the good example and not the three bad ones. Over time, after seeing different examples from different annotators the network will learn to hedge its bets and output plausible diverse predictions.”
Rakic is advised by Professors John Guttag of MIT EECS and Adrian Dalca of Harvard Medical School. She reports that both are incredibly supportive advisors and share a great sense of humor.
“[Prof. Dalca] brings a lot of creativity to my projects and helps me to aim for elegant and simple solutions,” Rakic said. “[Prof. Guttag] guides the project, making sure that there is a clear goal. He likes to ask — ‘What does success mean here?’ It is important for the delivery to be as good as the content. They help me work on my writing and presentation skills. And of course, there are the famous elevator pitches that we practice at lab meetings.”
Rakic noted that with AI everyone can see that things are moving extremely fast, and it’s hard to predict what things will look like, even in a year from now. Following her time at MIT, Rakic said she is unsure of her career plans.
“Right now, I’m just starting a new project, still looking into this idea of uncertainty. After that, I will think about what is next for me depending on where the field is at,” Rakic said. “I really enjoy research; the creativity and problem solving that it entails. It is also important for me to have an impact. That’s why I think that medical imaging is a great PhD research area for me.”
She added, “I would love to work in a stimulating environment where I can use what I learned at MIT, but also keep learning.”
You can find more information about Marianne Rakic and her research on her personal website: https://mariannerakic.github.io/.