Machine learning is a reality. It helps people understand their data and use this data to make effective decisions.
- Russ Greiner, Professor of Computing Science, University of Alberta
Brain tumours are insidious experts of invasion. They begin innocuously, just a single cell that has become chemically off-balanced. The cell multiplies, creating copies of its abnormal self, until it is a gluttonous mass threatening the brain’s ability to function normally. The most severe type of brain tumour can kill a person in just one year.
Brain tumours are often treated with radiation therapy. One approach is for doctors to detect where the tumour is by looking at MRIs (magnetic resonance images) of the brain. They then use radiation to kill the tumour cells.
Unfortunately, not all tumour cells in the MRIs are visible to the naked eye, not even the highly trained eyes of doctors. The standard procedure, therefore, is to irradiate both the visible portion of the tumour and everything else within a two-centimetre radius. While this procedure is often effective, it is far from perfect—tumour cells frequently lurk outside the two-centimetre radius. Also, this procedure kills any healthy brain cells in the region along with tumour cells.
Dr. Albert Murtha, a radiation oncologist at the Cross Cancer Institute, was wrestling with these problems, wondering how doctors could more accurately locate tumour cells and avoid killing healthy cells, when he heard about the work of Russ Greiner, a professor of computing science at the University of Alberta (U of A). Greiner’s specialty is machine learning. Murtha and Greiner began collaborating, and the result is the Brain Tumour Analysis Project (BTAP), which uses learning algorithms in an effort to better detect where brain tumours are and to predict where they will grow.
The BTAP team has created a computer program that can detect brain tumours all by itself, without the help of humans. The program, which is called an automated segmentation program (ASP), is based on a learning algorithm that has learned how to find tumours.
To master its job, the learning algorithm had to “train,” Greiner explains. In the training phase, the algorithm was given “labelled data”; these were hundreds of MRIs of brains with tumours in which doctors had already manually identified where the tumours were. The algorithm used the labelled data to determine methods for distinguishing tumour cells from normal cells. It then uses these methods to find tumours in the MRIs of new patients.
With the help of the learning algorithm, ASP can do a job that previously could only be done by rigorously trained human doctors. While the program isn’t perfect, neither are humans, Greiner points out. Doctors are busy, and they get tired. In an interview with Innovation Alberta, Murtha says, “(Analyzing MRIs to detect tumours) is a… time-consuming process. Having it automated has the advantage of speeding it up… And the faster we can be, the more efficient we can be and treat more patients in an efficient manner.”
Now that the project team has built an effective program for detecting tumours, the next challenge is using learning algorithms to predict how and where a tumour will grow once it is established in the brain. Eventually the project will also use learning algorithms to determine the type of tumours; for example, is a tumour Grade I, the least dangerous type, or is it Grade IV, the most dangerous type?
Although training a learning algorithm with labelled data is kind of like teaching a human through repetition, artificial intelligence and human intelligence are far from the same thing. Greiner says computers and humans each have their own special talents when it comes to intelligence.
Computers, for example, have superior memory, and they can compute huge math problems with lightning speed.
Humans, on the other hand, can think in analogies, a tough task for a computer. Humans can also easily analyze and respond to their environment when performing a task.
Maysam Heydari, a master’s student working on BTAP, gives the example of driving a car. Driving is an effortless activity for most humans, because we automatically know what to pay attention to, what to ignore, and how to adjust our driving to what’s going on around us. Driving a car is a much more challenging prospect for a computer, says Heydari, because a computer doesn’t have the human knack for focusing attention and reacting appropriately to the surrounding environment. Nonetheless, computers are getting better at driving cars.
“My own personal position is, I just don’t care (about the differences between artificial intelligence and human intelligence),” says Greiner. In his view, what’s important is that artificial intelligence helps human endeavours and augments human intelligence. Take the brain tumour project. “If we succeed here, we can reduce suffering and save lives,” says Greiner.
More about the Brain Tumor Analysis Project >