Enhancing surgical object detection in laparoscopic cholecystectomy with explicit positional relationship modeling

Laparoscopic Cholecystectomy (LC) is one of the most performed complex surgeries. Integrating Artificial Intelligence (AI) into LC shows great potential for assisting in anatomical structure detection. To be dependable, AI must be accurate, robust, and effective. In this study, a relation-based model was proposed to enhance surgical object detection in LC images. The model employs a positional relation encoder and refines progressive attention mechanism to analyze object relationships. Two widely used LC datasets were selected to validate the proposed model. We strictly followed the official split and evaluator protocols for fair comparison with benchmark models. The Macroscopic Correlation (MC) results revealed distinct differences in position relation strength between the two datasets, enabling comprehensive evaluation of the proposed models under different circumstances. The experimental results demonstrated the accuracy and effectiveness of the proposed models in both datasets. The proposed model outperformed the best-performing benchmark model by an improvement of 33.95 % in overall mean Average Precision (AP) on the Endoscapes dataset. For classes Cystic Plate and HC Triangle, the detection AP was improved by 90.32 % and 92.46 %, respectively. For the m2cai16-tool-locations dataset, the proposed models also demonstrated effective performance, improving the overall mAP by up to 17.97 % compared to benchmark models. The experimental results proved the accuracy and effectiveness of the proposed model. Due to the analysis of position relation, the detection of key objects is significantly improved. The postprocessing steps effectively reduce redundant bounding boxes by over 90 %. Future work could focus on expanding to more clinical and practical applications.