Latent Graph Representations for Critical View of Safety Assessment

Area of Science:

• Computer Vision
• Surgical Analytics
• Medical Imaging

Background:

• Assessing the critical view of safety (CVS) in laparoscopic cholecystectomy is crucial for patient safety.
• Existing methods often rely on semantic segmentation, which requires expensive annotations and can fail if segmentation is inaccurate.

Purpose of the Study:

• To develop a robust method for CVS prediction that overcomes the limitations of segmentation-based approaches.
• To reduce the annotation cost associated with training surgical safety assessment models.

Main Methods:

• Representing surgical images using a disentangled latent scene graph encoding semantic and visual features.
• Employing graph neural networks to process these scene graphs for anatomy-driven reasoning.
• Training the model using bounding box annotations with an auxiliary image reconstruction objective.

Main Results:

• The proposed method outperforms baseline methods when trained with bounding box annotations.
• The approach demonstrates effective scaling and maintains state-of-the-art performance when trained with segmentation masks.
• The graph representation enhances anatomy-driven reasoning and robustness to semantic errors.

Conclusions:

• The novel graph-based method offers a more efficient and robust approach to predicting critical views of safety in laparoscopic surgery.
• Reducing annotation requirements through bounding boxes makes advanced surgical safety assessment more accessible.
• This work advances the field of computer-assisted surgery by improving anatomical reasoning and generalization capabilities.