Page 50 - Read Online

Goyal et al. Art Int Surg. 2025;5:450-64 https://dx.doi.org/10.20517/ais.2024.108 Page 454

[7]
methods [15-17] . Consequently, one additional study was included in our analysis . The detailed reasons for
exclusion and the study selection process are illustrated in the PRISMA flowchart [Figure 1], created using
the online tool developed by Haddaway et al. [18,19] .

Study characteristics
This systematic review included six studies published between 2022 and 2024, focusing on laparoscopic or
robotic-assisted IHR procedures. The studies were conducted in three countries: the United States (n = 2),
France (n = 2), and Greece (n = 1). The hernia repair techniques investigated across the included studies
comprised laparoscopic total extraperitoneal (TEP), transabdominal preperitoneal (TAPP), and robotic-
assisted inguinal hernia repair (RIHR). Four studies utilized laparoscopic approaches [7,8,10,20] , while two
employed robotic-assisted techniques [9,21] .

The sample sizes varied significantly across the studies. One multicenter study, conducted in the USA,
[10]
reported a total of 619 cases, comprising 270 unilateral and 349 bilateral hernias . In contrast, the single-
center studies mostly included unilateral IHR, except for one study by Takeuchi et al. (2022), which
[8]
reported 86 unilateral and 33 bilateral cases . The periods of video data collection, when provided, spanned
from July 2019 to December 2022. However, two studies did not specify the time frame for their video
collections [7,21] . All studies used an AI model architecture to detect the surgical phase recognition in patients
undergoing minimal invasive IHR [Table 1].

Results on individual studies
Number of videos analyzed
From the included studies in the systematic review, encompassing both laparoscopic and robotic surgery
videos. The total number of videos analyzed varied across studies, ranging from 25 videos in the study by
Zygomalas et al. (2024) to 619 videos in the study by Ortenzi et al. (2023) . Among these, the studies
[7]
[10]
employed varying proportions of videos for specific tasks such as training, validation, and testing. A
majority of the studies prioritized training datasets, with proportions ranging from 75% to over 85% of the
total videos. Internal validation and testing were inconsistently reported, with only a few studies detailing
their allocation strategies [Table 2].

Number of surgeons involved
[10]
The number of surgeons contributing to the studies varied. Ortenzi et al.’s study (2023) had the largest
cohort of surgeons involved (n = 85), while Zygomalas et al.’s study (2024) reported the smallest group
[7]
[9]
[20]
(n =5). The studies by Takeuchi et al. (2023) and Zang et al. (2023) involved 11 and 8 surgeons,
respectively. Similarly, Choksi et al. (2023) included 8 surgeons in their study.
[21]
Pre-processing of videos
Prior to annotation, surgical videos underwent systematic pre-processing to ensure uniformity, reduce
computational load, and optimize datasets for analysis. Ortenzi et al. [10] utilized a Surgical Intelligence
platform (Theator Inc., Palo Alto, CA) to automatically capture, de-identify, and upload videos to a secure
cloud infrastructure. Takeuchi et al. [20] sampled videos uniformly over time to prevent selection bias, with
approximately 18.12 ± 11.59 images extracted per video, resulting in 2,356 training images. Additionally, 10
test images per video (a total of 300) were selected to represent the CVMPO visual inspection phase. Choksi
et al. down-sampled the videos to 1 frame per second (FPS) to reduce memory usage, which has been
[21]
[9]
shown to be sufficient for phase segmentation tasks. Similarly, Zang et al. processed videos using the
FFmpeg library to reduce the frame rate from 30 FPS to 1 FPS, a reduction that had minimal impact on
[7]
phase recognition performance. Zygomalas et al. extracted 1,095 images from 25 TAPP videos using
Bounding Boxes for object labeling, resulting in a dataset of 800 images for training and 295 for testing