Page 77 - Read Online

Novotny et al. Art Int Surg 2024;4:376-86 https://dx.doi.org/10.20517/ais.2024.52 Page 380

Traditionally, training is carried out on live animals, but this is costly and not always feasible. VR-based
[29]
microsurgery simulators, on the other hand, allow for cost-effective and realistic training . Through ML,
the skills and knowledge of surgeons can be compared and evaluated against countless others. A
statistically-based assessment can then be used to make training recommendations or suggestions for
improvement . Medical professionals engage in continuous learning throughout their careers, but
[30]
surgeons do not usually have the time to spend hours reading and evaluating publications and new
literature. AI can quickly analyze large amounts of new literature and provide the most important and
significant results in a summarized form . The ability to compare skills with surgeons around the world
[31]
and learn from other techniques and findings offers a whole new opportunity for global shared progress.

AI also enables the objective analysis of each surgeon’s results, offering individualized training programs
and suggestions for improvement. To be prepared for this emerging technology, a certain basic knowledge
of mathematics, computer science, ethical considerations and risks should already be included in the
[32]
university education of every medical professional . The basic prerequisite for this is the promotion and
approval of this technology in the training curriculum of every surgeon.

AI, ML AND DEEP LEARNING
ML enables computers to recognize patterns and learn from them. These are not explicitly programmed for
a specific use, but the program itself structures and categorizes the input using algorithms. The distinction
between supervised and unsupervised learning depends entirely on the intended goals. Supervised learning
should be used to work toward a specific goal or direction, as to specify certain guidelines externally.
Decision trees can be used to clearly define the criteria for analysis. Unsupervised learning is mainly used to
[33]
analyze large amounts of data for patterns . Another category of ML applications is where a program is
given a task to perform. The program simulates and tries all possible scenarios until the task is completed,
learning from its own successes and failures . ML algorithms can be used to predict the results of aesthetic
[34]
procedures, such as rhinoplasty or breast augmentation . Another advantage of ML is that multiple ML
[35]
programs can be combined and work together to make predictions and decisions that are far superior to
normal statistical models in terms of accuracy and precision .
[36]

ARTIFICIAL NEURAL NETWORKS
Artificial neural networks (ANNs) are systems based on the biological nervous system, which can be seen as
a subset of ML and the basis of all AI. In ANNs, the input is evaluated by the smallest units and distributed
to other units with different weights. The more input and information, the more accurate and reliable the
result . In terms of prediction rates, ANNs, sometimes in combination with MLs, have been able to
[26]
outperform traditional risk assessment tools with impressive superiority . ANNs are utilized in planning
[37]
body contouring procedures, such as liposuction, by analyzing patient-specific data and predicting how fat
can be removed or redistributed for optimal aesthetic results. The network helps ensure that the contours
[35]
created are as smooth and symmetrical as possible .

DEEP LEARNING
Deep learning (DL) has established itself as a leading method in ML, particularly for image pattern
recognition. It adds many different layers between the input and output layers, creating a complex network
of connections. It also incorporates hidden layers that significantly influence the result. It allows the
combination of supervised and unsupervised algorithms to solve large, complex problems and data sets . It
[38]
uses deep convolutional neural networks (DCNNs) to automatically learn data representations through
multi-layer neural networks. DCNNs extract relevant features from the training data by adjusting their
weights using backpropagation. No manually designed features are required. With sufficiently large and