Page 34 - Read Online

Talwar et al. Art Int Surg. 2025;5:418-24 https://dx.doi.org/10.20517/ais.2025.22 Page 422

Table 1. Models for preoperative surgical planning
Study Model type Task Training data Key findings
Love et al. Manual Predict need for AMR during 342 CT scans (1 axial slice per Rectus width to hernia width ratio inversely
[5]
2021 computation Rives-Stoppa hernia repair patient); all images used for correlated with need for AMR
statistical analysis
Elhage et Image-based Predict surgical complexity ~7,400 CT images Conventional neural network outperformed
[6]
al. 2021 deep learning and outcomes in abdominal surgeons in predicting surgical complexity;
wall reconstruction struggled with predicting rare complications
Ayuso et GANomaly deep Predict surgical complexity ~8,000 CT images Improved sensitivity (25% to 68% mesh
al. 2023 [9] learning and outcomes in AWR infection, 27% to 73% pulmonary failure);
precision declined
Wilson et Image-based Predict hernia recurrence; ~152 composite images (each Clinical-only model outperformed others;
[10
al. 2024 deep learning image-only, clinical data-only, created by averaging multiple image models underperformed due to small,
]
and mixed models axial slices containing the unbalanced dataset
hernia)
AMR: Additional myofascial release; CT: computed tomography; GANomaly: generative adversarial network anomaly; AWR: abdominal wall
reconstruction.

Another study explored the use of mobile thermal imaging and deep learning to detect SSIs following
cesarean section surgery in Rwanda, aiming to develop an accessible postoperative monitoring tool for
resource-limited settings . Two CNN models, a naïve CNN (baseline) and ResNet50 (pre-trained for
[16]
feature extraction), were trained on 500 thermal wound images, which were manually labeled by a
[16]
physician . These images were captured using a mobile thermal imaging device, which records heat
patterns emitted from the wound site. The naïve CNN achieved an accuracy of 0.86, while ResNet50
demonstrated an accuracy of 0.94 . The study was limited by the small number of SSI-positive images,
[16]
which may have affected the generalizability of the model. Expanding the dataset could improve model
performance and advance the development of a cost-effective postoperative monitoring modality.

CHALLENGES AND LIMITATIONS
AI-based imaging shows promise in VHR but faces barriers to widespread clinical adoption [17,18] . A Lancet
editorial highlights the need for high-quality datasets to minimize bias to prevent innovations from
[19]
exacerbating healthcare disparities in resource-limited settings . These populations may have unique
physical characteristics that may be inaccurately assessed by the AI technologies.

Secondly, as AI explodes in complexity, deciphering the “black box,” referring to a model’s opaque decision-
making process, has become exponentially harder [19-21] . Physicians who rely upon AI-based tools may have
difficulty justifying decisions, improving models, and gaining new insights . Numerous explainable AI
[21]
models have been proposed in medicine; however, current approaches for image-analysis models report
explanations that are too complex to be understood by a user [22,23] .

Furthermore, the computational costs restrict further adoption . Despite falling digital technology costs,
[24]
the World Health Organization notes that access remains inequitable . Thus, the lack of digital
[25]
infrastructure may limit the use of AI in resource-poor settings . Deep learning systems rely on high-
[25]
performance GPUs, a dependency that is becoming increasingly costly as the demand for computing power
continues to rise .
[26]
FUTURE WORK
Despite the promising application of AI in VHR, challenges must be addressed to ensure equitable and
clinically reliable AI implementation.