A team of computer scientists from the University of Alberta, Canada, and the derivative company MEDO, have worked on new research that has allowed the development of a new deep learning model capable of learning to identify diseases from medical scans in a very more accurate and faster .
As we have already explained to you on different occasions, deep learning consists of a type of machine learning , within artificial intelligence (AI) ; specifically, computer algorithms that find patterns in large data sets, then produce models that can later be used to make predictions.
Precisely, these models work best when they learn from hundreds of thousands or even millions of examples. But, in the field of medical diagnostics, it is true that the system presents a unique challenge, as researchers typically only have access to a few hundred medical scan and test images for privacy reasons.
This raises quite a few problems, since “when you train a deep learning model with so few instances, its performance tends to be poor.” To try to solve this, the researchers tried to address how to develop effective models of deep learning for medical tasks with few instances of training, taking advantage of the knowledge present in the medical literature with which to better guide the learning process of AI .
Thus, while the improvements in performance come from the training of the algorithm from medical images, the researchers also used the “probabilistic” diagnosis, provided indirectly by medical specialists, allowing the algorithm to learn the patterns that characterize each disease in scans, thereby predicting which disease appears on a new patient’s examination, or whether it appears originally healthy .
It is, as the researchers indicate, “to provide radiologists with better tools that make their work easier, faster and more effective.” Especially if we take into account that there is a “problem of shortage of medical personnel, exacerbated in developing countries”.
The truth is that the development of this type of model could be very useful in allowing medical specialists to make better decisions .
As the scientists indicate, this research includes important contributions from Pouneh Gorji, a posthumous graduate of the Department of Computer Science, who died in the tragedy of flight PS752 with his wife, in January 2020, when they were traveling to Iran to marry together with others. four members of the community of the aforementioned Faculty of Sciences (176 people died in the accident).
Therefore, for the team, this investigation becomes an opportunity to honor Gorji’s legacy and work.
Reference: Sample Efficient Learning of Image-Based Diagnostic Classifiers Using Probabilistic Labels. arXiv: 2102.06164v1
