Insert Figure 1
Over the past years, the ML subfield of deep learning has gained
tremendous popularity, as it has yielded superior results in analyzingunstructured data such as medical images, text data, and audio
data. This technique is based on large artificial neural networks
(ANNs). ANNs form networks inspired by the biological animal brain,
consisting of multiple layers of processing units called neurons. Deep
learning methods can detect complex data relationships by automatically
compressing data and distilling relevant features in various levels of
abstraction. This makes it different from statistical approaches such as
regression methods, which require explicitly defining independent
variables and making assumptions about their relationship to the outcome
variable. Another advantage of deep learning is its ability to continue
learning and improve performance with larger datasets. Besides
applications in computer vision, which is the ability to interpret image
data by an AI system, deep learning has also propelled natural language
processing (NLP) forward, which is the capacity of a computer to
understand written and spoken human language. We refer to recent,
extensive reviews that cover the subfields of deep learning and its
applications in medicine. A state-of-the-art example is the detection of
diabetic retinopathy from retinal images, for which the IDx-DR
deep-learning-based software has been FDA-approved and validated in a
clinical setting. Relatedly, deep learning approaches have outperformed
trained physicians in breast cancer detection using imaging data, with
currently nine applications FDA-approved. Some of the latest AI
breakthroughs that ignited the general public’s interest are based on
deep learning approaches. These involve generative models that are
trained to create new data. Examples include deep fakes, DALL-E (an
OpenAI application that creates figures and art based on written
descriptions), and most recently, ChatGPT, an AI tool that generates
highly realistic written text based on user prompts. Figure S1 displays
two AI-generated illustrations of this review’s topic.
Learning strategies
A useful categorization of AI is made on the learning strategy, which
defines how an algorithm learns from data. Three different approaches
are distinguished: supervised, unsupervised, and reinforcement learning.
We provide a conceptual framework to structure AI applications based on
learning strategy, learning goal, data modality, and medical domain in
Figure 2.