Table 3.

A summary of the articles describing the role of AI applied to LC

Authors	Year of publication	Dataset and characteristics of the methods	Algorithm and AI
Mascagni et al. [37]	2020	78 endoscopic videos of consecutive LC procedures performed at the Digestive Surgery Department of the Nouvel Hospital Civil	The inter-observer agreement was calculated with Cohen’s kappa; doublet view method and the binary scheme
Mascagni et al. [38]	2022	2854 images from 201 LC videos were annotated and 402 images were segmented	The inter-observer agreement was calculated with Cohen’s kappa
Madani et al. [39]	2022	308 anonymized LC videos from 37 countries (including all continents), 153 surgeons and 136 different institutions AI predictions were evaluated using tenfold cross-validation against annotations by expert surgeons. A tenfold cross-validation technique evaluated the performance of GoNoGoNet and CholeNet	ffmpeg 4.1 software extracted frames from videos; Deep Convolutional Neural Network (CNN; ResNet50); Pyramid Scene Parsing Network (PSPNet) was used for pixel-wise semantic segmentation GoNoGoNet AI and CholeNet
Laplante et al. [40]	2023	308 anonymized videos from 37 countries, 153 surgeons and 136 different institutions High-volume expert surgeons delineated the boundaries of the Go and No-Go zones (semantic segmentation)	ffmpeg 4.1 software extracted frames from videos; Think Like A Surgeon software was used by expert surgeons to delineate the boundaries of the Go and No-Go zones Visual Concordance Test (VCT) calculated the expert consensus of annotated pixels for Go and No-Go zones GoNoGoNet AI
Khalid et al. [6]	2023	The database was composed of 308 anonymized LC videos originating from 37 countries across 153 surgeons and 136 different institutions. Two groups: BDI group and Control group. 11 LC videos of BDI group was annotated by GoNoGoNet and compared to another 11 LC videos with cholecystitis of control group	Convolutional Deep Neural Network (DNN); U-Net architecture (the optimized version of the GoNoGoNet algorithm);
Endo et al. [41]	2023	230 videos of LC conducted from 2019 to 2021 in a single center; 95 cases that remained on video with mild inflammation to be noted; cases with severe inflammation and abnormal biliary anatomy were excluded 1754 images from Calot triangle dissection scenes, 1610 images of LM-EHBD, 1503 images of LM-EHBD, 1623 images of LM-S4, and 1505 images of LM-RS Test datasets: 19 cases, comprising 190 images of LM-EHBD, 186 images of LM-EHBD, 192 images of LM-S4, and 190 images of LM-RS Questionnaire survey was given to beginners and experts surgeons	YOLOv3
Nakanuma et al. [42]	2023	10 cases of LC performed in a single center to evaluate the feasibility of LC using an intraoperative AI-assisted landmark recognition The landmark detection system was connected to the endoscopic integrated operating room (EndoALPHA; Olympus Corp.) The method for quantitatively evaluating the accuracy of landmark detection by AI was the DICE coefficient. The DICE coefficient is the reference value for detecting the accuracy of AI and shows the degree of agreement between expert surgeons for the annotations of the anatomic regions of interest The Kruskal–Wallis test was used to evaluate the DICE coefficient	YOLOv3 connected with the endoscopic integrated operating room (EndoALPHA; Olympus Corp.)
Fujinaga et al. [43]	2023	20 cases of LC performed in a single center External evaluation committee (EEC) evaluated the adequacy of the detection times of the landmarks detected using AI 4-point rubric questionnaire was used to evaluate the accuracy of landmark detection and the contribution of cross-AI in preventing BDI	Cross-AI system for landmark detection and surgical phase recognition: an endoscope (OLYMPUS LTF-S190-10; Olympus Corp., Tokyo, Japan), a video processor (VISERA ELITE II; Olympus Corp.), and a desktop computer that had two graphics processing units (Quadro RTX 6000 and Quadro RTX 4000; NVIDIA Corp., Santa Clara, CA, USA)
Kawamura et al. [36]	2023	72 LC videos, and 23,793 images were used for training data, performed in a single center	EfficientNet-B5: a CNN to classify images and perform prediction; Sharpness-Aware Minimization: a SAM an optimizer that perfects the learning parameters of EfficientNet-B5 to smooth the information derived from the diversified images for a label
Golany et al.[44]	2022	371 LC videos from 4 hospitals and data set Cholec80: 294 videos were used to train the AI model, 77 videos were used to test the AI model. Each video was divided into 10 steps by 2 expert surgeons Experienced surgeons noted the phases, adverse events and CSV of LCs The inter-rater agreement score was used to confirm the quality of the annotations	MS-TCN—Multi-Stage Temporal Convolution Network. is a set of temporal convolution layers that capture temporal connections. The final layer of the MS-TCN gives the prediction of the surgical stage for each frame of the LC video Resnet50 independently classifies and predicts the surgical phase. For each frame (input), Resnet50 produces a numerical vector of visual features. The set of input vectors are combined to form a sequence of feature vectors representing the entire LC video and fed into the MS-TCN model
Cheng et al. [48]	2022	Dataset: 163 LC videos collected from four medical centers. 90 videos were labeled by expert surgeons. 63 LC videos were used to test the AI model The mean concordance correlation coefficient was used to compare the reliability between the two surgeons	Super Video Converter software (version 1.0); The Anvil Video Annotation Research Tool software was used to annotate the videos; FFmpeg software was used to extract frames from videos; Convolutional neural network (CNN) captured the spatial characteristics; Long Short-Term Memory (LSTM) analyzed the temporal information Anaconda (Anaconda, Inc, Austin, TX) in Python 3.6.5. NVIDIA (NVIDIA, Santa Clara, CA) Tesla V100 graphics processing unit (GPU) trained the visual and temporal models