NUL STUDENT IS BUILDING A ROBOTIC, ACCURATE EARLY CANCER DETECTOR

The Artificial Intelligence (AI) machine is being created by Mahao Molise, the National University of Lesotho (NUL) student. It seeks to replace skilled radiologists with highly trained computers that detect cancer cells with a better degree of accuracy.

When you train machines to think like humans; that is called artificial intelligence. “Through coding, we give the machine some brain-like capabilities to make them intelligent and identify cancer cells,” he said.

(Catch him here: +266 5971 3551) 

This young fellow, who is supervised by the indefatigable Messrs Napo Mosola and Kopano Moeketsi, is beaming with energy as he describes his intelligent system. “It is a known fact that even the most highly trained and well-meaning humans make mistakes when it comes to identifying cancer,” he said as he introduced a system that would be part of the NUL Innovation Hub.

As he set his eyes on the development of this robotic system, he was once invited to a cancer conference in Lesotho that would further reinforce his commitment to this solution. The conference which was attended by first ladies from across Africa was called “12 th Stop Cervical, Breast and Prostate Cancer Conference.”

In it, he said, “I listened to Nigeria’s first lady detailing how she was once diagnosed with “cancer” in Nigeria only to discover that she had no such thing in the United Kingdom.” Another gentleman was sent from Lesotho to India for “prostate cancer” only for the Indians to confirm that, actually, his problem was skin cancer.

In literature, he found there was always a possibility that even the best trained humans who read image scans for cancer identification could get it wrong sometimes.

To reduce the error, his system seeks to go beyond just the image analysis. It asks for the assistance of the artificial intelligence and big data.

“I combine computer vision (image processing) with deep learning (part of the artificial intelligence) to help me identify cancer cells and their kind more accurately,” he said.

It starts with turning a computer into something like a human brain [don’t get us wrong, the human brain is still the most sophisticated machine ever seen in the entire universe—unless this is the case of a human brain praising itself]. As you know, human brains have neural networks that make humans intelligent and give them ability to learn.

So his job, he said, is to create similar neural networks, “not in human brains but inside a computer.” The neural networks are created through some form of sophisticated coding which we won’t dare reveal here lest we literally blow our minds.

Once the neural networks have been created, they are trained, “in the same manner you train humans to identify some things.” For instance, he said, how do you train humans to identify a house where there are so many forms of houses? In Lesotho, you have “polata,” “heise,” “mokhoro,” “apostetse,” and so on. How do you know that they are all houses when they are clearly different?

“It is because they have certain features which are all common to them,” Mahao said. “It could be windows, it could be doors, it could be bricks, it could be roofing.”

In the same way, “we trace a particular form of cancer and train the computer to study common features even if the cancer can always come in slightly different forms.”

But how on earth do you train a computer? It is just as complicated as it sounds. Back to our house example, for a human to really know what a house is, despite a house coming in different forms, he has to see a lot of houses in his lifetime. That is what learning means.

In the same way, his neural networks have to “see” so many pictures of cancer cells, for a cancer of a certain type, that they end up “knowing” and identifying those cells. The more pictures the better. We call those numerous pictures big data.

For instance, to train his machine, Mahao has in his possession, a humongous 50 Gigabytes worth of pictures of cancer cells! And, “for the perfection we are looking at, this massive data is still not enough, but it is good nevertheless,” he said.

Once the computers are trained, now the detection part.

Say you get an image of a personal's lung that went through a CT scan. Before you apply artificial intelligence to it, you simply pass it though the computer’s image processing software where you clean it. “You will have to remove a lot of other things captured by the image which may includes fat, water, etc, to leave a lung tissue.”

Once these images are clean, they are passed through the already trained neural network and something interesting happens. “Unlike trained humans,” he said, “who look at the whole picture,” the neural networks “look” at the picture pixel by pixel until it identifies cancer cells if they exist at all in the lung.

Analysing an image pixel by pixel can detect even the smallest cancer cells which would otherwise be impossible to detect.