NCSA’s Data Analytics Team Takes On Cancer

If you watch a television show based in a hospital, it can seem like diagnosing cancer is often a relatively simple process. Some aspects of these shows aren’t far off the mark. Tumors can be seen on scans and even some x-rays, and they can sometimes be felt during a physical examination. Anyone who’s had a lump or bump removed knows the next step is a biopsy. But this is where Hollywood often skips a few steps to keep the plot moving, often triumphantly diagnosing the disease within the day. Even when it’s simple to find a tumor, it’s not always easy to diagnose.

Determining if a tumor is benign or cancerous is the first step, but doctors can’t always look at cells under a microscope and know it’s cancer. Some cancer cells, for example, low-grade cancer cells, can look just like normal cells, making certain types of cancer diagnosis a time-consuming process.

Once a sample is determined to be cancer, the next step is determining what type of cancer the doctors are dealing with. There are over 200 types of cancer, and they are grouped by where they start. For example, knowing a patient has breast cancer will guide an oncologist toward a type of treatment, but they need more information to make an effective treatment plan. Not all breast cancer patients need the same treatment. Some need aggressive chemotherapy, while others can be treated with radiation alone. Cancer treatments are hard on the body, so doctors try to avoid overtreating if possible. Undertreatment is also dangerous. Needless to say, finding the most accurate diagnosis can go a long way to help treat cancer with the best results.

Xiaoxia Liao, the technical program manager in NCSA’s Data Analytics team, has been working with her team to find ways to make diagnosing breast cancer simpler and more accurate. Her team has created a bioimage analysis technique that utilizes deep learning to train supercomputers to predict patient prognosis using images of tumor samples.

Using deep learning speeds up the process of diagnosis. A computer can analyze an image down to the pixel, and it can do it with a speed and accuracy not possible with human eyes. This is important because cancer can be made up of different types of cells – cells with a range of genetic and molecular properties – and the differences between normal cells and cancer cells are microscopic. Another hurdle facing doctors is that high-end cancer diagnosis tools that could aid in determining the specifics of a cancer cell aren’t available in every hospital. With a bioimage analysis program, a doctor could theoretically submit images of tumor samples to an organization that diagnosed cancer and have an answer in minutes instead of physically shipping the sample to a specialized cancer center to be analyzed.

I am passionate about applying my software engineering experience and data science knowledge into problem-solving. I learned a lot from working with bioimaging and deep learning experts in this project. My team and I look forward to utilizing NCSA high-performance computers and collaborating with researchers on more interesting scientific and engineering projects in the future.

Xiaoxia Liao, technical program manager, NCSA Data Analytics

The work that Liao’s team has done so far has already been helpful for researchers. Liao’s work was included in a Nature article where researchers showed that using these bioimage analysis techniques could improve treatment accuracy. The study shows that using these techniques can dramatically reduce the overtreatment and undertreatment of breast cancer, especially in patients with smaller tumors.

Building on this success, the data analytics team is already expanding their work into other fields, proposing to use their techniques to predict the success rate of newborn livestock using images of livestock embryos.