Lung nodule interpretation in computer tomograhpy scans

How does a computer scientist make a machine learn?

"We provide the computer with a truth; then, it investigates the relationship between that truth and available medical data - in our current research, computer tomography images of lung nodules - to predict truth in other instances," says Jacob Furst, associate professor, School of Computing.

"In other words, in this research, we're asking the computer to make connections, to see patterns not immediately apparent to humans because these patterns are embedded in millions of pixels," he explains. "A computer will never replace a radiologist. But if a computer could 'read' a CT scan, the radiologist could use it to be more efficient and more certain in interpreting images."

Daniela Raicu, associate professor, School of Computing, and Furst were working on systems for recognizing abdominal tissues when the National Institutes of Health released a data set of 149 lung nodules, challenging computer scientists to develop an application that could "read" the images. The data set also includes interpretations of each image from four expert radiologists.

Could a computer accurately predict the label that a radiologist would apply?

"If you ask two radiologists the meaning of a single image, you're likely to get two different responses. This variability increases complexity," says Raicu, explaining one of the challenges in the research.

"A computer program - like the human radiologist - will interpret the image with some degree of uncertainty. When the computer suggests an answer, it will say, in effect, 'This nodule is diseased; my confidence is 85 percent.' The system would not say 'yes' or 'no' absolutely - the real-world is not that black-and-white. On the other hand, a computer's 'opinion' could bridge the gap between dissenting radiologists."

The subtlety of diagnostic medicine is not the only challenge in the research, as Raicu explains:

"We don't have well-defined benchmarks, and everyone building computer-aided diagnostic systems has that same problem. If the computer makes a decision, is that decision indeed valid? Because DePaul doesn't have a medical school of its own, we can't confirm that data with biopsies.

"Also, we're working with a relatively small data sample of only 149 cases. How do we know what the critical data set size must be to ensure that we capture enough information about the 'truth' so we can generalize the process to work well enough on new cases? These are some hard problems, which we are unlikely to solve in their entirety within the next decade or two. This research is that cutting-edge!"

Furst and Raicu are working with two other researchers outside the university: a radiologist at Northwestern Memorial Hospital and a medical physicist at University of Chicago. During the summer, funding from the National Science Foundation allows 10 undergraduates from around the country to participate in this research. During the academic year, the team includes four Ph.D. students, one M.S. student and one undergraduate.

"In six years, we've worked with probably 100 students," says Furst. "For this summer, we received 90 applicants for the 10 spots. So, students see this project as a great way to gain some research experience and prepare for graduate school."

Raicu agrees: "I teach this research in my undergraduate classes - it's a great way to convey theory to students. We're pushing them to think about what they're learning in a real-world way."