Page 14 - Read Online

Page 259 Collins et al. Art Int Surg. 2025;5:254-69 https://dx.doi.org/10.20517/ais.2024.92

that the model was not originally trained on. Monitoring also helps identify issues like incorrect data
labeling or shifts in regulatory requirements, prompting updates to the model or its application to ensure
ongoing reliability and compliance.

Although we have outlined the primary steps involved in AI model development, a detailed discussion falls
beyond the scope of this article. For a comprehensive examination of AI model development in healthcare,
we refer readers to a selection of foundational texts, including [31,32] .

RESEARCH IN AI AND AWS
Researchers from various other institutions are advancing AI technologies specifically to enhance AWS,
requiring close interdisciplinary collaborations among surgeons, AI researchers, software engineers, and
clinical researchers. Drawing on our own experience, this article explores how advancements in AI
technologies can enhance AWS, organized by five key application areas: (1) preoperative planning and
outcome prediction; (2) intraoperative assistance; (3) emergency surgery; (4) postoperative care; and (5)
surgery education. This section reviews the literature, divided into the above five categories.

Preoperative planning and outcome prediction
Computed tomography (CT) is generally considered the most effective imaging modality for accurately
assessing the size and location of abdominal hernias, rectus diastasis, and any associated muscle atrophy, as
well as the hernia’s proportion relative to the abdominal wall. This detailed information can guide surgeons
in selecting the most appropriate surgical approach [open vs. minimally invasive surgery (MIS)],
determining the optimal positioning and fixation of meshes, and assessing the need for additional
interventions, such as botulinum toxin injections, preoperative pneumoperitoneum, or component
separation techniques.

However, despite their importance, these findings are often missing in CT reports, as radiologists may not
be familiar with evaluating the abdominal wall or understanding the specific details required for surgical
planning. For this reason, surgeons often do not consult CT images, relying instead on physical examination
and patient history.

The underutilization of preoperative CT has been recognized as a limitation of current practice, driving
recent efforts to improve radiology hernia reports for surgeons . AI has great potential to accelerate this
[33]
[34]
process. An important advance is the concept of a 3D virtual clone (or “digital twin”) , where AI models
(typically deep neural networks - DNNs ) reconstruct 3D models of a patient’s specific anatomy from a
[35]
[26]
preoperative image, with a process known as image segmentation. Neural networks such as nnUNet can
delineate a wide variety of tissues, and recently, TotalSegmentator , adapted from nnUNet, delineates 117
[36]
anatomical and pathological classes in CT images using a training dataset of 1,228 patients. Digital twin
research has matured with products approved for medical use. Visible Patient, a startup founded in 2013
[34]
that spun off from IRCAD France, provides an online service to create patient-specific anatomy models
from CT and MR images. This combines DNNs and human oversight for output correction and validation.

There are several preliminary works to generate automatic patient-specific 3D models of hernias [37,38] . Zhang
et al. developed a method to differentiate abdominal wall tissue from the hernia sac to improve 3D hernia
visualization and measurement and assist treatment planning . This was a small pilot study; however,
[37]
mature models for finer-grained differentiation of muscles, fascia, hernia sacs, and vessels will soon exist,
trained on larger cohorts with robust multi-centric external validation.