Demystifying Artificial Neural Networks
In part 2 of our series on the basics of digital technology, we explore the deep learning tools that can improve medical image analysis and much more.
By John Halamka, M.D., President, Mayo Clinic Platform and Paul Cerrato, MA, senior research analyst and communications specialist, Mayo Clinic Platform
Just how important are artificial neural networks (ANNs) in healthcare? Given the fact that they are now being used in algorithms that help diagnose several diseases, enhance medical image analysis, and improve medical record summarization, it’s fair to say that lacking a basic knowledge of ANNs will put clinicians at a significant disadvantage in today’s healthcare ecosystem.
ANNs get their name from the neural networks that exist in the brain. Torsten Wiesel and David Hubel won a Nobel Prize for their work in deciphering how the brain interprets images, including the ability of simple neurons to understand the orientation of an image’s edges, in all 360 degrees of orientation. These simple neurons then send this information to complex neurons that can sense various shapes; as Figure 1 illustrates, all this information eventually enables the brain to construct an abstraction—a cat in this example.
This collection of simple and complex neurons is similar to the layers in an ANN, but with the ANN, the input data might include thousands of images from a radiology data set or photos of the retina or the skin to help identify diabetic retinopathy or melanoma, respectively. The neurons in the brain are similar to the nodes in the ANN, which exist in hidden layers, as Figure 2 illustrates. As we pointed out in an earlier column, each node is assigned a numerical weight, which corresponds to the strength of the feature in the layer. A higher number indicates that the feature—an irregular shape for a suspected skin cancer, for instance—means it is more likely to suggest the presence of the malignancy, while a lower number assigned to a smooth round circumference suggests a normal mole.
Figure 2
Unfortunately, when ANNs initially perform these calculations, they make many mistakes. The network uses a process called backpropagation to fix them. It looks back at the mistakes to help the program readjust the algorithm, changing the numerical values of its weights, thereby improving its accuracy or predictive power.
Among the many types of ANNs that have surfaced over the years, convolutional neural networks (CNNs) have proven especially useful and are frequently employed to assist with image recognition. The term convolution can confuse readers without a math background. In everyday English, something is convoluted if it’s very complex and difficult to follow. In informatics, it refers to a “mathematical operation that combines two functions to describe the overlap between them. Convolution takes two functions and “slides” one of them over the other, multiplying the function values at each point where they overlap, and adding up the products to create a new function.” In plain English, the mention of sliding helps demystify the process. CNNs act like stencils that recognize portions of an image when they pass over them. When the filter detects a matching element in the image—an A in the name Ada Lovelace in Figure 3—the convolution process generates a strong image, which is then mapped to a feature map.
Figure 3
(Source: Cerrato, P, Halamka J. Reinventing Clinical Decision Support. Talor & Francis, 2020.)
This filtering process enables CNNs to detect complex patterns, which is why they have proven valuable in medical image analysis, helping to improve the interpretation of X-rays and MRI scans.
Once data is processed through convolutional layers, it has to pass through three additional steps: activation functions, pooling layers, and fully connected layers, which are beyond the scope of this series of primers. For a more detailed explanation of how CNNs work, JAMA has created an easy-to-understand video on the topic.
In part 3 of our series of primers, we’ll take a closer look at the technology behind large language models.
Recent Posts
By John Halamka and Paul Cerrato— Guidelines and guardrails for the safe use of AI require more than regulation. The Coalition for Health AI has created a framework to reduce these risks and improve the safety and effectiveness of these models.
By John Halamka and Paul Cerrato — The capabilities of generative AI continue to grow. Using them wisely will likely improve clinical decision making.
By John Halamka and Paul Cerrato — The Oxford Dictionary says disease is a disorder of structure or function, especially one that has a known cause and a distinctive group of symptoms, signs, or anatomical changes. It’s time to rethink that simplistic definition.