A systematic review of annotation for surgical process model analysis in minimally invasive surgery based on video

NeurostimML: a machine learning model for predicting neurostimulation-induced tissue damage

Objective. The safe delivery of electrical current to neural tissue depends on many factors, yet previous methods for predicting tissue damage rely on only a few stimulation parameters. Here, we report the development of a machine learning approach that could lead to a more reliable method for predicting electrical stimulation-induced tissue damage by incorporating additional stimulation parameters.Approach. A literature search was conducted to build an initial database of tissue response information after electrical stimulation, categorized as either damaging or non-damaging. Subsequently, we used ordinal encoding and random forest for feature selection, and investigated four machine learning models for classification: Logistic Regression, K-nearest Neighbor, Random Forest, and Multilayer Perceptron. Finally, we compared the results of these models against the accuracy of the Shannon equation.Main Results. We compiled a database with 387 unique stimulation parameter combinations collected from 58 independent studies conducted over a period of 47 years, with 195 (51%) categorized as non-damaging and 190 (49%) categorized as damaging. The features selected for building our model with a Random Forest algorithm were: waveform shape, geometric surface area, pulse width, frequency, pulse amplitude, charge per phase, charge density, current density, duty cycle, daily stimulation duration, daily number of pulses delivered, and daily accumulated charge. The Shannon equation yielded an accuracy of 63.9% using akvalue of 1.79. In contrast, the Random Forest algorithm was able to robustly predict whether a set of stimulation parameters was classified as damaging or non-damaging with an accuracy of 88.3%.Significance. This novel Random Forest model can facilitate more informed decision making in the selection of neuromodulation parameters for both research studies and clinical practice. This study represents the first approach to use machine learning in the prediction of stimulation-induced neural tissue damage, and lays the groundwork for neurostimulation driven by machine learning models.

Implementation of artificial intelligence-based computer vision model in laparoscopic appendectomy: validation, reliability, and clinical correlation

BACKGROUND

Application of artificial intelligence (AI) in general surgery is evolving. Real-world implementation of an AI-based computer-vision model in laparoscopic appendectomy (LA) is presented. We aimed to evaluate (1) its accuracy in complexity grading and safety adherence, (2) clinical correlation to outcomes.

METHODS

A retrospective single-center study of 499 consecutive LA videos, captured and analyzed by 'Surgical Intelligence Platform,' Theator Inc. (9/2020-5/2022). Two expert surgeons viewed all videos and manually graded complexity and safety adherence. Automated annotations were compared to surgeons' assessments. Inter-surgeons' agreements were measured. Since 7/2021 videos were linked to patients' admission numbers. Data retrieval from medical records was performed (n = 365). Outcomes were compared between high and low complexity grades.

RESULTS

Low and high complexity grades comprised 74.8 and 25.2% of 499 videos. Surgeons' agreements were high (76.9-94.4%, kappa 0.77/0.91; p < 0.001) for all annotated complexity grades. Surgeons' agreements were also high (96.0-99.8%, kappa 0.78/0.87; p < 0.001) for full safety adherence, whereas agreement was moderate in partial safety adherence and none (32.8-58.8%). Inter-surgeons' agreements were high for complexity grading (kappa 0.86, p < 0.001) and safety adherence (kappa 0.88, p < 0.001). Comparing high to low grade complexity, preoperative clinical features were similar, except larger appendix diameter on imaging (13.4 ± 4.4 vs. 10.5 ± 3.0 mm, p < 0.001). Intraoperative outcomes were significantly higher (p < 0.001), including time to achieve critical view of safety (29.6, IQR 19.1-41.6 vs. 13.7, IQR 8.5-21.1 min), operative duration (45.3, IQR 37.7-65.2 vs. 25.0, IQR 18.3-32.7 min), and intraoperative events (39.4% vs. 5.9%). Postoperative outcomes (7.4% vs. 9.2%) including surgical complications, mortality, and readmissions were comparable (p = 0.6), except length of stay (4, IQR 2-5.5 vs. 1, IQR 1-2 days; p < 0.001).

CONCLUSION

The model accurately assesses complexity grading and full safety achievement. It can serve to predict operative time and intraoperative course, whereas no clinical correlation was found regarding postoperative outcomes. Further studies are needed.

Artificial Intelligence in Cataract Surgery: A Systematic Review

Purpose

The purpose of this study was to assess the current use and reliability of artificial intelligence (AI)-based algorithms for analyzing cataract surgery videos.

Methods

A systematic review of the literature about intra-operative analysis of cataract surgery videos with machine learning techniques was performed. Cataract diagnosis and detection algorithms were excluded. Resulting algorithms were compared, descriptively analyzed, and metrics summarized or visually reported. The reproducibility and reliability of the methods and results were assessed using a modified version of the Medical Image Computing and Computer-Assisted (MICCAI) checklist.

Results

Thirty-eight of the 550 screened studies were included, 20 addressed the challenge of instrument detection or tracking, 9 focused on phase discrimination, and 8 predicted skill and complications. Instrument detection achieves an area under the receiver operator characteristic curve (ROC AUC) between 0.976 and 0.998, instrument tracking an mAP between 0.685 and 0.929, phase recognition an ROC AUC between 0.773 and 0.990, and complications or surgical skill performs with an ROC AUC between 0.570 and 0.970.

Conclusions

The studies showed a wide variation in quality and pose a challenge regarding replication due to a small number of public datasets (none for manual small incision cataract surgery) and seldom published source code. There is no standard for reported outcome metrics and validation of the models on external datasets is rare making comparisons difficult. The data suggests that tracking of instruments and phase detection work well but surgical skill and complication recognition remains a challenge for deep learning.

Translational Relevance

This overview of cataract surgery analysis with AI models provides translational value for improving training of the clinician by identifying successes and challenges.

Implementation of Artificial Intelligence-Based Computer Vision Model for Sleeve Gastrectomy: Experience in One Tertiary Center

INTRODUCTION

Sleeve gastrectomy (SG) is the most common metabolic and bariatric procedure performed. Leveraging artificial intelligence (AI) for automated real-time data structuring and annotations of surgical videos has immense potential of clinical applications. This study presents initial real-world implementation of AI-based computer vision model in sleeve gastrectomy (SG) and external validation of accuracy of safety milestone annotations.

METHODS

A retrospective single-center study of 49 consecutive SG videos was captured and analyzed by the AI platform (December 2020-August 2023). A bariatric surgeon viewed all videos and assessed safety milestones adherence, compared to the AI annotations. Patients' data were retrieved from the bariatric unit registry.

RESULTS

SG total duration was 47.5 min (interquartile range 36-64). Main steps included preparation (12.2%), dissection of the greater curvature (30.8%), gastric transection (28.5%), specimen extraction (7.2%), and final inspection (14.4%). Out of body time comprised 6.9% of the total video. Safety milestones components and AI-surgeon agreements included the following: bougie insertion (100%), distance from pylorus ≥ 2 cm (100%), parallel to lesser curvature (98%), fundus mobilization (100%), and distance from esophagus ≥ 1 cm (true-100%, false-13.6%; kappa coefficient 0.2, p = 0.006). Intraoperative complications included notable hemorrhage (n = 4) and parenchymal injury (n = 1).

CONCLUSIONS

The AI model provides a fully automated SG video analysis. Outcomes suggest its accuracy in four of five safety milestone annotations. This data is valuable, as it reflects objective performance measures which can help us improve the surgical quality and efficiency of SG. Larger cohorts will enable SG standardization and clinical correlations with outcomes, aiming to improve patients' safety.

NeurostimML: A machine learning model for predicting neurostimulation-induced tissue damage

Objective

The safe delivery of electrical current to neural tissue depends on many factors, yet previous methods for predicting tissue damage rely on only a few stimulation parameters. Here, we report the development of a machine learning approach that could lead to a more reliable method for predicting electrical stimulation-induced tissue damage by incorporating additional stimulation parameters.

Approach

A literature search was conducted to build an initial database of tissue response information after electrical stimulation, categorized as either damaging or non-damaging. Subsequently, we used ordinal encoding and random forest for feature selection, and investigated four machine learning models for classification: Logistic Regression, K-nearest Neighbor, Random Forest, and Multilayer Perceptron. Finally, we compared the results of these models against the accuracy of the Shannon equation.

Main Results

We compiled a database with 387 unique stimulation parameter combinations collected from 58 independent studies conducted over a period of 47 years, with 195 (51%) categorized as non-damaging and 190 (49%) categorized as damaging. The features selected for building our model with a Random Forest algorithm were: waveform shape, geometric surface area, pulse width, frequency, pulse amplitude, charge per phase, charge density, current density, duty cycle, daily stimulation duration, daily number of pulses delivered, and daily accumulated charge. The Shannon equation yielded an accuracy of 63.9% using a k value of 1.79. In contrast, the Random Forest algorithm was able to robustly predict whether a set of stimulation parameters was classified as damaging or non-damaging with an accuracy of 88.3%.

Significance

This novel Random Forest model can facilitate more informed decision making in the selection of neuromodulation parameters for both research studies and clinical practice. This study represents the first approach to use machine learning in the prediction of stimulation-induced neural tissue damage, and lays the groundwork for neurostimulation driven by machine learning models.