Machine learning could help increase reliability of cancer diagnosis

July 26, 2016

After years of research and persisting issues with the reliability of cancer imaging techniques, accurate and confident cancer diagnosis is still a significant hurdle for clinicians and patients.

This challenge leads to a large number of painful, invasive and often unnecessary biopsies, according to Dr. Mehdi Moradi, a 2013-2014 Peter Wall Early Career Scholar who is working to reduce this added distress for patients undergoing the cancer diagnosis process.

He says clinicians and researchers tend to err on the side of caution, assuming anything on an image that is slightly suspicious may be cancer and require a biopsy to confirm.

“Why? Because the cost of missing cancer is supposedly higher than the cost of sending a patient to an unnecessary biopsy,” says Dr. Moradi, an Assistant Professor with UBC’s Department of Electrical and Computer Engineering who holds an engineering degree from University of Tehran and a PhD from Queen’s University in biomedical computing.

“But on the other hand, if the number is thousands and thousands of unnecessary biopsies– which is the case for breast cancer– then it starts to not make a lot of sense.”

Cancer imaging techniques, such as MRIs (above), PET scans and ultrasounds do not provide adequate diagnostic reliability, meaning doctors must often send their patients on for painful biopsies, says Dr. Moradi.

According to Dr. Moradi, more than 80 per cent of women sent for the invasive biopsies used to detect breast cancer end up with a negative result.

In the case of prostate cancer, 50 per cent of the men who go home with a negative biopsy result end up back for more testing, either because the procedure failed to accurately sample the tumour, or perhaps sampled a low grade tumour while there is a more serious high grade tumour in a different area. Imaging deficiencies often make it difficult to biopsy the exact part of the tissue affected. Patients may return home, only to end up sent back by doctors for another painful biopsy after other prostate cancer indicators– such as a rising Prostate Specific Antigen– continue to become stronger.

Many of these men eventually end up with positive cancer diagnosis, says Dr. Moradi.

“It’s a very frustrating situation. If you can reduce the number of repeated biopsies, obviously a lot of pain is removed.”

For Dr. Moradi, the answer to increasing the reliability of our imaging techniques could lie within machine learning.

Machine learning is a branch of artificial intelligence that looks to construct and study computer systems that can “learn” from large amounts of data.

In the social sciences, for example, machine learning could allow researchers to determine the likelihood an Internet user will engage with a certain product, video or advertisement, based on their user profile.

“There are a lot of parameters that identify the profile of that person – how old they are, if they are male or female, where he or she lives, and their previous behaviour on the Internet”, says Dr. Moradi. Machine learning looks at the interaction of these factors to determine a probable outcome.

“On the imaging side, we are realizing that no single parameter is good enough to detect cancer,” he says.

In the medical imaging field, machine learning makes it possible to analyze multiple sources of information at once, as opposed to a radiologist looking through separate images from PET scans, MRI, and ultrasounds. There is simply too much information, and much information that is not visible, says Dr. Moradi, for radiologists to analyze these different information sources and determine a confident diagnosis.

Researchers can use machine learning to construct a patient profile based on the information gathered from various imaging techniques, then correlate this information with test results from tissue post-biopsy.

“If you do this long enough with many patients, you can gradually build a machine learning model that not only learns to find cancer, but can also predict the aggressiveness of the cancer.”

“The idea is basically to improve our own confidence in the imaging methods by combining several modalities and by building a data-driven machine learning approach, that over time can learn from the data and predict the presence and grade of cancer,” Dr. Moradi says.

Since Dr. Moradi’s model improves with time and more information, his current research goals centre around collecting more data from patients and cancer research centres. He eventually hopes that the technology could mean more confidence in imaging diagnosis and a better quality of life for those going through the process. 

He says his time as an Early Career Scholar with the Institute has reinforced this focus, giving him a sense of belonging to a community that is working towards different problems, but shares the goal of solving some of today’s most significant issues.

He calls his work a mission rather than simply a career.

“We’ve had a lot of advances in a lot of areas of science”, says Dr. Moradi, who lost his grandfather to compilations of prostate cancer.

“But if you look at the statistics of mortality caused by cancer, we haven’t done very well as a society in the last 20 to 30 years.”

“We are still really in a situation where we are struggling to reduce the number of deaths caused by cancer… all of this is motivation.”