2025 ISAKOS Biennial Congress ePoster
The Reliability Of Convolutional Neural Network Models In Classifying The Shoulder Arthroscopic Field's Visual Clarity
Son Quang Tran, MD, Cần Thơ VIETNAM
Thanathep Tanpowpong, MD, Bangkok THAILAND
Thun Itthipanichpong, MD, Bangkok THAILAND
Danaithep Limskul, MD, Bangkok THAILAND
Napatpong Thamrongskulsiri, MD, Bangkok THAILAND
Bao Nguyen Tu Thai, MD, Can Tho VIETNAM
Minh Cong Bui, MD, Can Tho VIETNAM
Department of Orthopedics, Faculty of Medicine, Chulalongkorn University, Bangkok, THAILAND
FDA Status Not Applicable
Summary
We developed CNN models for classifying the visual clarity of shoulder arthroscopic images and videos, with DenseNet169 demonstrating the highest reliability and agreement with rater assessments, making it a potential objective tool for future research.
ePosters will be available shortly before Congress
Abstract
Background
The assessment of visual clarity in the shoulder arthroscopic field by raters tends to be highly subjective. We aim to develop CNN models for the classification of the visual clarity of arthroscopic shoulder images and then to evaluate the reliability and agreement of these models with the assessment of raters.
Methods
We retrospectively reviewed intraoperative videos from 113 patients who underwent shoulder arthroscopic surgery from August 2020 to August 2024. Videos from 63 patients (group T) were used to create a dataset for training the models. Videos from the remaining 50 patients (group R) were utilized to evaluate the reliability and agreement of the trained models. Images extracted from the videos were assessed for visual clarity with a 3-grade scale. Subsequently, we implemented transfer learning techniques for the pre-trained CNN models involving DensetNet169, DenseNet201, Xception, InceptionResetV2, VGG16, and ViT. The ultimate trained models were utilized to predict the visual clarity of images in the independent datasets generated from videos in group R. The agreement and reliability of the trained predictive models compared with raters in classifying the visual clarity of shoulder arthroscopic images were reported with percentage of agreement and weighted kappa coefficients. Similarly, the agreement and reliability of the trained predictive models compared with raters in classifying the visual clarity of shoulder arthroscopic videos were evaluated using bias, limits of agreement (LoA), and intra-class coefficients.
Results
Almost all of these models performed well in the validation step, with an accuracy of 90% and above. The trained predictive models involving InceptionResNetV2, DenseNet169, and ViT exhibited a high percentage of agreement at 95.6%, 94.6%, and 94.4%, respectively, while their weighted kappa coefficients were reported to exceed 0.8. The trained DenseNet169 models demonstrated the highest reliability in the evaluation of the visual clarity of whole surgical videos with an ICC of 0.9, a bias of 0.027, and the narrowest LoA.
Conclusion
The trained DenseNet169 predictive model was reliable enough to be utilized as an objective measure of the visual clarity of the shoulder arthroscopic field for further research.