Page 33 - Read Online

Page 214 Boutros et al. Art Int Surg 2022;2:213-23 https://dx.doi.org/10.20517/ais.2022.32

roots in mathematics, statistics, neuroscience, linguistics, and even philosophy, AI has evolved from
programs that took advantage of rule-based systems based on expert knowledge to automated feature
[1]
extraction through deep learning . The rapid growth and evolution of AI approaches can leave one feeling
overwhelmed and confused about how these technologies will impact hepatopancreaticobiliary (HPB)
surgery, the obstacles to its clinical translation, and the role that HPB surgeons can play in accelerating AI’s
development and ultimate clinical impact. Thus, we review some of the basic terminology and current
approaches in surgical AI and how HPB surgeons can influence its future in surgery.

MACHINE LEARNING
Maier-Hein et al. define surgical data science (SDS) as the capture, organization, analysis, and modeling of
[2]
data sets with the aim of improving interventional healthcare . AI, more specifically, machine learning, has
played a significant role in the advancement of surgical data science, though it is important to note that
machine learning is not the only technique used in SDS. To better understand AI work that has been
performed thus far in HPB surgery, it is important to recognize that there are several learning strategies
within machine learning. Two popular approaches are classical machine learning, where humans
specifically guide algorithms to include specific features in their analyses, and neural networks, where
features are automatically derived by the algorithm. Neural networks are designed to function in a manner
akin to biological neural networks - a series of computational units (i.e., neurons) are arranged in layers and
process input data to yield a target output. Deep learning refers to neural networks that are arranged in
greater than 3 layers (i.e., input layer, > 1 hidden layer, output layer). The use of many hidden layers enables
the network to learn much more complex relationships than would be possible through a simple 3-layer
neural network. Each neuron within the network receives input data, processes it using a mathematical
function, and yields an output. Each neuron may be weighted differently depending on the overall
performance of the network, and these weights are tuned over multiple examples to generate optimal
performance on a given task. Deep learning is a technique that uses layers of computation to extract
information from an input in progressive fashion until a desired output is produced . Progressive
[3]
extraction refers to different levels of information. For example, initial layers in a deep learning network
may extract such things as edges in an image, while subsequent layers may utilize information on those
edges to identify faces and then specific people.

Common learning strategies that clinicians are likely to encounter in the AI literature are supervised,
unsupervised, and semi-supervised learning. Supervised learning refers to task-driven learning wherein an
[4]
algorithm learns from data that has been annotated by humans . For example, a researcher may have
annotated a dataset of the critical view of safety in hundreds of photographs from laparoscopic
cholecystectomy. An algorithm would “learn” from these photographs and annotations and subsequently
attempt to label new images from laparoscopic cholecystectomies with the critical view of safety. While
supervised learning through neural networks removes the effort needed to select features manually for
analysis, it has several challenges. Large amounts of data are typically needed to optimize the performance
of neural networks, especially deep learning. Furthermore, the data must be annotated to appropriately
identify the target output. That is, if one wishes to accurately identify photos of the critical view of safety in
cholecystectomy, a dataset composed primarily of neurosurgical photographs may not have much value. In
supervised learning, the labels on which the algorithms are trained are a critical part of the data.

Unsupervised learning, on the other hand, does not rely on labels to learn phenomena. Instead,
unsupervised learning analyzes the patterns inherent in the data to determine whether identifiable groups
are present. Stated perhaps more simply, unsupervised learning allows the data to be clustered into groups
with similar features, and a human can interpret those clusters to determine whether features may identify