Page 12 - Read Online

Page 257 Collins et al. Art Int Surg. 2025;5:254-69 https://dx.doi.org/10.20517/ais.2024.92

networks (ANNs), which are broadly inspired by the structure and function of the human brain.

ANNs consist of artificial neurons arranged in layers. Each neuron has input connections to receive data
and output connections to transmit processed information to subsequent neurons. Every connection
between neurons has a weight, which determines the strength of the information passed into the neuron.
These weights are crucial to the network’s performance, and during training, the weights are automatically
adjusted to maximize the model’s performance according to a performance metric, which is selected
according to the task at hand, such as anatomical structure recognition accuracy.

A single artificial neuron processes its inputs according to simple mathematics; in general, the inputs are
multiplied by their weights, then the weights inputs are added together, and then an activation function is
applied, which determines whether the neuron should activate or remain inactive based on the input it
receives. The activation function plays a crucial role in allowing neural networks to learn and make
decisions by introducing non-linearity to the output. Various activation functions exist. For example, the
sigmoid activation function maps the input to a value between 0 and 1, acting like a smooth “on/off” switch
that controls the strength of a neuron’s output. This makes it useful for tasks like binary classification, where
the output can represent probabilities.

While a single artificial neuron has limited practical use, the complexity and capability of an ANN arise by
using many interconnections between many artificial neurons. In particular, increasing the depth of the
network (i.e., the number of layers of neurons) allows ANNs to learn complex patterns from training data,
[19]
[20]
enabling tasks such as medical image analysis , language translation , and generative AI applications [21,22]
[24]
like image, video , and speech creation. The size of contemporary ANNs varies significantly, primarily
[23]
depending on the range of tasks they are designed to perform - broader task ranges typically require larger
ANNs. The largest ANNs used for understanding and generating natural language are large language
models (LLMs), such as OpenAI’s GPT-4 (used in ChatGPT ). While OpenAI has not released official
[25]
statistics, some open-source alternative LLMs - such as Meta’s Lama 405B - report having 405 billion
trainable parameters. In contrast, AI models intended for use as medical devices tend to be smaller due to
narrower intended use. For instance, models used to delineate structures in radiology images, such as
nnUNet , typically have tens of millions of trainable parameters, requiring substantially less time,
[26]
computational resources, and financial investment for model training.

The rapid growth of AI, ML, and DL in recent years is attributed to several key factors. First, the explosion
of data in the “big data” era has provided essential training material for advanced DL models. Second,
significant advancements in computational power - fueled by powerful GPUs and specialized hardware -
have enabled faster processing and more efficient model training. Third, the emergence of open-source,
[27]
[28]
sophisticated programming frameworks like TensorFlow and PyTorch has made it easier for
researchers and AI engineers to train complex models. Lastly, increased investment in AI research and
development, alongside growing interest from both the public and private sectors, has accelerated progress
in these fields.

AI model development for healthcare applications
AI development for healthcare applications typically progresses through five main phases. The first phase,
data collection and model training, involves gathering high-quality, relevant data to teach the AI model. A
critical aspect of this phase is data annotation, where input-output pairs are labeled or “tagged” with specific
information, enabling the model to learn in a process known as supervised learning. This process is similar
to how a student studies for an exam by practicing with questions, looking at the correct answers, and then