According to new research by a team of University of Alberta computing scientists and the U of A spinoff company MEDO, a new deep-learning model can learn to identify diseases from medical scans faster and more accurately. The breakthrough model was developed by a team of researchers from the Faculty of Science, including contributions from Pouneh Gorji, a graduate student who died in Flight PS752.
Deep learning is a subfield of artificial intelligence that is a type of machine learning. Deep learning techniques are computer algorithms that find patterns in large sets of data, producing models that can then be used to make predictions. These models perform best when they are trained on hundreds of thousands, if not millions, of examples. However, the field of medical diagnostics presents a unique challenge in that researchers typically only have access to a few hundred medical scan images due to privacy concerns.
“When a deep-learning model is trained with so few examples, its performance tends to be poor,” said Roberto Vega, the study’s lead author and a graduate student in the Department of Computing Science.
A new deep-learning model can learn to identify diseases from medical scans faster and more accurately, according to new research by a team of University of Alberta computing scientists and the U of A spinoff company MEDO.
“In our study, we looked at how to learn effective deep-learning models for medical tasks with few training instances. The main idea is that we can use the knowledge in medical literature to better guide the learning process.”
The algorithm’s performance is improved by training it on medical images as well as their corresponding “probabilistic” diagnosis, which is provided indirectly by medical experts. This method enables the algorithm to learn the patterns that characterize each disease in scans, allowing it to predict what disease is shown in a new patient’s scan or whether the scan appears healthy.
“Our approach improved the model’s classification accuracy while also providing meaningful confidence in its prediction, providing an estimate of the probability that a disease is present in a scan,” Vega said.
“We hope that by conducting this research, we will be able to provide radiologists with better tools that will make their jobs easier, faster, and more effective. We also have a shortage of medical personnel, which is exacerbated in developing countries. Our hope is that we will be able to create models that will allow our medical experts to make better decisions.”
The work of many
Pouneh Gorji, a posthumous graduate of the Department of Computing Science and a victim of the Flight PS752 tragedy, made significant contributions to the research. Gorji and Arash Pourzarabi returned to Iran in January 2020 to marry, and were two of four members of the Faculty of Science community among the 176 people killed on Flight PS752.
The research is also an opportunity for the team to honor Gorji’s legacy and her critical contributions to the project. Vega explained that when the researchers started the project, the algorithm was solely focused on hip dysplasia. Gorji was working on developing machine learning models for identifying fatty liver when she joined the team, and it was her presence on the team that resulted in a redesign of the algorithm and a breakthrough in performance.
“Because the original algorithm did not work for fatty liver, Pouneh and I began collaborating to solve the problem. We discovered an important flaw in the original approach after several weeks and were able to propose a solution—one that involved a new mathematical way of tackling a section of the model, “Vega stated.
“The collaborative work we did was critical to the success of our approach, and it is what led to the eventual publication of this research. This publication would not exist without her help.”
The paper, titled “Sample Efficient Learning of Image-Based Diagnostic Classifiers Using Probabilistic Labels,” will be presented at the 21st International Conference on Artificial Intelligence and Statistics (AISTATS). It’s also available on arXiv.