June: Machine Learning


Tricia Chinnery, BSc
Contact: tchinne@uwo.ca
June 8, 2021

A 63-year-old male walks into the London Regional Cancer Program (LRCP) at Victoria Hospital and presents with a 4 cm tumour in his lungs. He is diagnosed with stage 3 lung cancer and is prescribed radiation therapy to treat the tumour. He receives a Computed Tomography (CT) scan (which can be thought of as a 3D X-ray) as part of the routine clinical workflow, and he begins receiving daily radiation treatments. However, apart from the size and location of the tumour, there are limited clues available to help physicians determine if the treatment will be successful or not. Therefore, there is a need for computer software to help physicians determine if the patient will be cancer-free after treatment is complete. If there was a software tool that could predict if a patient’s cancer will return, it could guide physicians in altering a patient’s course of treatment. For example, a physician could prescribe a higher radiation dose to patients more likely to have their cancer return.

Medical imaging, such as CT, Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET), are commonly used to diagnose and treat disease. With advancements in medical image analysis, it is now possible to extract numbers from standard medical images. This numerical data could not only help provide information that is not visible to physicians’ eyes but help predict outcomes for patients as well. At the Gerald C. Baines Centre located in Victoria Hospital, the collaborative vision of a team led by Drs. Ward, Mattonen, and Palma is to achieve positive changes in the treatment of patients living with cancer through improvements in imaging technology and diagnostics. Their imaging research laboratories focus on a bench-to-bedside approach, and they develop computer software to correlate imaging features with clinical outcomes. Their software may assist physicians in clinical decision-making, ultimately improving outcomes for cancer patients.

There is a large amount of hidden data in medical images that goes unused, and since cancer patients get scanned upon beginning treatment anyways, extracting numbers from their scans to aid in predicting outcomes is very practical. It is a non-invasive, inexpensive way to find meaningful information that could save lives. When cancer patients receive CT scans in the clinic, their tumours are outlined on the image before receiving radiation treatment. Researchers in the Baines Laboratories are developing software to extract information that aims to describe tumour characteristics. They hypothesize that information such as the brightness in the CT image or the texture of the tumour may uncover details about the tumour’s behaviour. They also extract shape and size details, which give information about the volume and surface area of the tumour. Next, algorithms that rely on self-learning to find patterns within data will identify which pieces of information best predict which patients will be cured after treatment. But how are these algorithms created? And how are these models developed?

Machine learning finds patterns in data and uses them to make predictions. Machine learning involves teaching a computer to “learn” to do a task without being explicitly programmed to do that task. The figure below outlines the difference between normal computer software and machine learning software.

Normal software is a set of step-by-step instructions written by a programmer, intended to achieve a particular output. In other words, we are telling the computer what to do.

Machine learning software finds patterns on its own and tries to predict a certain output. It is software that writes software! We tell the computer how to figure out the answer itself, using the data we feed it. This “learning” process generally requires large amounts of data, and it is purely mathematical. To begin, we show the computer examples of patient images and information about their tumours and provide the computer with the “answers” (whether the patient’s cancer returned after treatment or not). Then, the computer finds a mathematical function that represents these inputs (tumour data) and outputs (answers) through a process called training. It is called this because we are “training” a computer to learn a relationship between the inputs and outputs. If the function (a.k.a. “model”) can learn from the data successfully, it will be able to make a prediction about a patient it has never seen before. The better the model learns from training data, the more accurately the model can come up with responses when asked about whether new patients starting treatment will be cured of their cancer.

The ability to self-learn makes machine learning very useful amongst cancer applications. Not only can these models be used to help predict if a patient’s cancer will return after treatment, but they may also be used to predict if the patient will develop side effects! In the example of our lung cancer patient, if the physician was able to know that the patient’s tumour would return after treatment through a machine learning model, the physician may have opted to prescribe a more aggressive treatment, such as a higher radiation dose to the tumour, or to add chemotherapy to the treatment plan. This adjustment may result in a higher chance of successful treatment for the lung cancer patient! These tools being developed in the Baines Laboratories are non-invasive and may lead to more personalized medicine options. Their goal is to improve outcomes and increase quality of life for all cancer patients.

To learn more about work being done in the Baines Laboratories, visit their website at https://www.bainesimaging.com/


Images from: “What Is Machine Learning? – Visual Explanations | Data Revenue.” [Online]. Available: https://www.datarevenue.com/en-blog/what-is-machine-learning-a-visual-explanation.

“How Does Machine Learning Work? – dummies.” [Online]. Available: https://www.dummies.com/programming/big-data/data-science/how-does-machine-learning-work/.

American Cancer Society, “Non-Small Cell Lung Cancer Stages.” [Online]. Available: https://www.cancer.org/cancer/lung-cancer/detection-diagnosis-staging/staging-nsclc.html.