Abstract
Efficient operating room (OR) scheduling is critical for hospital logistics, yet traditionally relies on subjective surgeon estimates or labor-intensive manual phase-tracking. This paper proposes an automated approach to predict remaining surgery duration by fusing endoscopic video with data from surgical devices. By employing self-supervised learning on unlabeled temporal sequences, the developed multimodal model achieves a 37% average error across diverse procedure types, significantly outperforming traditional baselines without requiring manual annotation.
Introduction
The core challenge in OR management is unpredictability. Surgical workflows vary significantly based on patient anatomy and unforeseen complications. While previous AI attempts at duration prediction have shown promise, they typically rely on "surgical phase detection," which requires thousands of hours of manually labeled video data. This creates a significant labeling bottleneck and limits the ability of models to generalize across different types of procedures.
To address these gaps, the authors developed a context-aware method that utilizes multimodal data already generated during laparoscopic procedures—such as video feeds and device telemetry—to predict progress without explicit human labeling.
Methods
The researchers proposed a suite of models based on Recurrent Convolutional Neural Networks (CNNs). The architecture utilizes a CNN to extract features from each frame and a Gated Recurrent Unit (GRU) to maintain a temporal history of the surgery.
Key components of the multimodal fusion include:
- Self-Supervised Vision (V-Net): To avoid manual labels, the vision model was pretrained using a Siamese approach. The network learned the temporal flow of surgery by determining the correct order of random frames, effectively learning what "progress" looks like.
- Tool Presence (T-Net): A ResNet-152 detects the presence of 12 surgical tools, serving as a proxy for specific surgical tasks.
- Device Data (D-Net): The model integrates 14 discrete signals from OR equipment, including insufflator gas flow/pressure, light intensity, and camera exposure settings.
- Multimodal Fusion (VTD-Net): The final model concatenates these streams into a single input for the GRU.
Results and Discussion
The study evaluated these models on two datasets: MultiType (80 diverse interventions) and Cholec80 (80 standardized gallbladder removals). The results demonstrate that while vision is the primary driver of accuracy, multimodal fusion provides critical context.
As shown in the results table, the full VTD-Net achieved an overall average error of 37% on the MultiType dataset. Accuracy improved as the surgery progressed; by the halftime mark, the error dropped to approximately 28%. In contrast, a "Naive" baseline of guessing the average duration for a surgery type resulted in a 57% error, highlighting the significant predictive value of the AI model.
On the standardized Cholec80 benchmark, the vision-tool model (VT-Net) achieved a 23% average error, outperforming previous state-of-the-art methods like Twinanda et al.’s RSDNet (25% error).
Conclusion
This research demonstrates that self-supervised, multimodal learning can effectively predict surgical duration without the need for expensive manual annotations. While the model performs best on standardized procedures, it shows robust improvements over traditional baselines even in diverse surgical environments. Future work may involve incorporating additional data streams, such as anesthesia telemetry, to further refine predictions and account for intraoperative complications.
Summary updated