Comparative Analysis and Exploration of 3DResNet-18 and InceptionV3-GRU with Temporal Segment Network (TSN) Framework for Laparoscopic Surgical Video Expertise Classification

Laparoscopic surgery is a widely adopted minimally invasive procedure that requires surgeons to master complex skills such as suturing, knot-tying, and needle-passing. Traditional assessment of these skills is often subjective and prone to bias, relying heavily on manual evaluation by expert surgeons, which can vary between evaluators. We applied deep learning models to automate surgical skill evaluation to address this issue and move towards a more objective and standardized assessment method. In this study, we utilized two advanced architectures—3D ResNet-18 and InceptionV3-GRU—within a Temporal Segment Network (TSN) framework to classify skill levels using the publicly available JHU-ISI Gesture and Skill Assessment Working Set (JIGSAWS) dataset. We focused on optimizing temporal sampling by adjusting the number of frames and frame intervals in the video data. Our findings show that capturing longer sequences of actions improved accuracy for suturing and needle-passing tasks while capturing more detailed motions enhanced performance for knot-tying. Our findings suggest that capturing longer sequences of actions improved accuracy for suturing and needle-passing tasks while capturing more detailed motions enhanced performance for knot-tying. The 3D ResNet-18 model achieved 100% accuracy across all tasks, significantly outperforming the InceptionV3-GRU model, which achieved 85.71% for suturing, 77.42% for knot-tying, and 100% for needle-passing. These results demonstrate the superior capability of the 3D ResNet-18 model in surgical skill classification and highlight the critical role of temporal optimization in improving performance across different surgical tasks.