AI-powered real-time detection and classification of peritoneal nodules in colorectal cancer using multi-center laparoscopic data: A feasibility study.

OBJECTIVE

To explore the feasibility of developing an intraoperative real-time detection model for peritoneal nodules in laparoscopic colorectal cancer surgery using computer vision and deep learning, and to provide a preliminary assessment of its potential for assisting surgeons in nodule identification.

METHODS

Laparoscopic colorectal cancer surgery videos from three regional general hospitals were collected. Systematic peritoneal exploration segments were extracted and converted into image frames to construct the dataset. Surgeons performed annotation of peritoneal nodule locations and benign/malignant classification based on pathological results. The YOLOv11l architecture was employed to train separate models for peritoneal nodule detection and benign/malignant classification. Model performance was evaluated on the test set using metrics such as mAP@50. A comparative experiment with novice surgeons was designed to assess the model's accuracy and real-time performance.

RESULTS

Based on a dataset constructed from 165 multi-center surgical videos, the YOLOv11l abdominal peritoneal nodule detection model developed by us performed well on the test set, with an AP@50 of 0.846 for nodule detection and an mAP@50 of 0.835 for benign and malignant classification. The real-time inference speed reached 14 FPS, meeting the requirements for real-time laparoscopic video analysis. Compared with mainstream detection models, YOLOv11l demonstrated better performance in both nodule detection and benign and malignant classification tasks. Human-machine comparison showed that the model's recognition time was significantly shorter than that of novice surgeons, while the recognition accuracy reached 89.47%.

CONCLUSION

This feasibility study demonstrates that a real-time peritoneal nodule detection and classification system based on YOLOv11l can be developed. The model shows promising accuracy and speed in identifying and classifying peritoneal nodules during laparoscopic colorectal cancer surgery. However, further validation with larger, patient-level datasets is needed before clinical application.