Artificial Intelligence for Workflow Analysis in Colorectal Surgery: A Multicentric, Cross-Procedural Development and Generalization Study
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - AI-ColoWorkflow automates phase and step recognition in minimally invasive colorectal surgery videos using a vision transformer and temporal convolutional network. Training on pooled multicenter, multi-procedure data generally improved phase-recognition generalization, though procedure-specific models remained stronger for some step-level tasks.
- The model combines fine-tuned DINOv3 frame features with a hierarchical multi-stage temporal convolutional network jointly trained for phase and step recognition.
- On held-out data, it achieved 73.01% macro F1 for phase recognition and 39.82% for step recognition.
- The pooled global model usually outperformed center- and procedure-specific alternatives, with the main exception being procedure-specific step recognition.
- Generalization phase-recognition F1 averaged 48.42%, supporting hybrid training strategies for finer-grained surgical workflow analysis.
Sources (1)
Artificial Intelligence for Workflow Analysis in Colorectal Surgery: A Multicentric, Cross-Procedural Development and Generalization Study
TL;DR - AI-ColoWorkflow automates phase and step recognition in minimally invasive colorectal surgery videos using a vision transformer and temporal convolutional network. Training on pooled multicenter, multi-procedure data generally improved phase-recognition generalization, though procedure-specific models remained stronger for some step-level tasks.
- The model combines fine-tuned DINOv3 frame features with a hierarchical multi-stage temporal convolutional network jointly trained for phase and step recognition.
- On held-out data, it achieved 73.01% macro F1 for phase recognition and 39.82% for step recognition.
- The pooled global model usually outperformed center- and procedure-specific alternatives, with the main exception being procedure-specific step recognition.
- Generalization phase-recognition F1 averaged 48.42%, supporting hybrid training strategies for finer-grained surgical workflow analysis.