ASA - Automated Assessment of Surgical Quality in Distal Gastrectomy: Development of a Novel Computer Vision Model Based on the Critical View of Quality (CVQ)

Back to 2026 Abstracts

Automated Assessment of Surgical Quality in Distal Gastrectomy: Development of a Novel Computer Vision Model Based on the Critical View of Quality (CVQ)
*Jeesun Kim^¹, *Dotan Asselmann^², *Tamir Wolf^², *Seong-Ho Kong^¹, *Do Joong Park^¹, *Hyuk-Joon Lee^¹, Gerald Fried^³, Han-Kwang Yang^⁴
^¹Seoul National University Hospital, Seoul; ^²Theator, Inc., Palo Alto, California; ^³McGill University, Montreal, Quebec; ^⁴National Cancer Center, Goyang

OBJECTIVES
Video-based assessment is emerging as a critical determinant of surgical quality. Yet, an objective intraoperative metric for gastric cancer surgery is lacking. To address this gap, we developed the Critical View of "Quality" (CVQ), a visual anatomical metric quantifying the completeness of dissection across five major lymph node (LN) stations in distal gastrectomy (DG). This study describes the integration of CVQ with a next-generation computer vision model designed to automatically recognize vascular landmarks, quantify lymphadenectomy quality, and ultimately link intraoperative performance to clinical outcomes.

METHODS
Two hundred minimally-invasive DG procedures were reviewed from a prospectively maintained clinical registry containing detailed preoperative variables and postoperative outcomes. Key vascular landmarks in LN#4sb, LN#6, LN#5/12a, suprapancreatic, and LN#1/3a stations were manually annotated and paired with CVQ scores (0-20 scale). Accompanying clinicopathologic variables and outcomes were incorporated to establish a comprehensive training set for model development.
For automated CVQ assessment, pre-computed video features were processed using a transformer-decoder architecture that applies cross-attention between learnable queries and temporal video representations (Figure 1). This design enables selective focus on time points demonstrating key vascular exposures and lymphadenectomy completeness, producing a compact feature embedding subsequently used for CVQ score prediction and component classification. This approach allows the model to handle variable-length dissections while maintaining anatomical specificity. The dataset is now being used to train and refine automated CVQ prediction, step segmentation, and quantitative assessment of lymphadenectomy quality.

RESULTS
CVQ scores demonstrated substantial variation (range 9-19), reflecting real-world differences in the clarity of vascular exposure across LN stations. Higher CVQ scores were associated with increased LN retrieval (r=0.23, p<0.05), supporting its validity as a measure of surgical radicality. Inter-rater agreement for CVQ component scoring was strongest in stations with consistent vascular exposure, such as LN#6. Early model prototypes accurately identified major vascular landmarks and demonstrated promising alignment with expert-assigned CVQ scores, indicating feasibility for full automation as training expands.

CONCLUSIONS
CVQ represents a rigorous, visually grounded metric capable of capturing the thoroughness of lymphadenectomy in DG. Its integration with a computer vision model provides a scalable pathway toward automated, anatomy-driven assessment of oncologic radicality. By linking intraoperative video features with clinical outcomes from a large institutional registry, this approach has the potential to define new standards for surgical quality assurance in gastric cancer and support future clinical translation of AI-based performance evaluation.

Figure 1. Computational architecture for automated CVQ scoring in distal gastrectomy

Back to 2026 Abstracts