Why it is good to use deep learning for medical image diagnosis – and even better to use probabilistic deep learning
How can artificial intelligence help to save lives?
The diagnosis based on histopathological images is considered as the gold standard for many different cancer types. The pathologist looks at the previously extracted tissue on microscope images and tries to characterize the cancer (if present) based on certain patterns. The correct diagnosis is crucial for the correct treatment of this life-threatening disease.
As pattern recognition is one of the strengths of deep learning algorithms, researchers hope to make this diagnostic process more accurate, faster and standardized with the help of artificial intelligence.
Let’s have a look at the theory of deep learning..
In a mathematical formulation we can describe this problem as follows: Given a histopathological image patch x and it’s corresponding (cancer) class y, we want to find a model that approximates the function f which maps x to y: y = f(x). This is done by minimizing the error between the true class y and the models prediction ŷ = mϕ(x) with respect to the models parameters ϕ. After training, these models are often able to provide accurate predictions. The problem with common deterministic deep learning models is that they are black boxes and the predictions are point estimates without confidence intervals. This makes it hard to know, when to trust in a prediction and when not.
What advantages offer probabilistic methods?
Probabilistic methods in comparison define a probability distribution over functions to approximate the underlying function f. This allows us to treat the prediction ŷ as a random variable with a probability distribution instead of a point estimation. The mean (or mode, depending on the problem) of this distribution describes the estimated class while the variance indicates the uncertainty of the prediction. The predictions with a small variance can therefore be considered as reliable. Another advantage is that probabilistic methods can be embedded in a probabilistic problem formulation. As pathologists have a different grade of experience and even expert pathologists can disagree in the diagnosis, we can model the true label y as a random variable to capture this uncertainty. These so called crowdsourcing models are one focus of the PhD of the ESR8 position.
And this is us:
The research group of Prof. Rafael Molina at the Department of Computer Science and Artificial Intelligence, Universidad de Granada, focuses on probabilistic methods like Gaussian Processes with Bayesian inference and their combination of deep learning. Recent applications for these techniques in histopathology were blind color devonvolution, cancer classification, crowdsourced WSI datasets and many more to come: current areas of research are multiple instance learning and active learning, just to mention a few.
We hope you enjoyed reading and thank you for your interest!
Arne Schmidt – ESR8