Predicting Anastomosis or Stump Leakage After Laparoscopic Gastrectomy: A Deep Learning Approach to Intraoperative...

Area of Science:

• Surgical Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Postoperative leakage is a significant complication following laparoscopic gastrectomy for gastric cancer.
• Early detection of leakage is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To develop and validate deep learning models for predicting postoperative leakage at anastomosis sites during laparoscopic gastrectomy.
• To assess the performance of various deep learning architectures in classifying normal versus leakage images.

Main Methods:

• Analysis of 10,256 laparoscopic images from 2,035 gastric cancer patients undergoing gastrectomy.
• Training six deep learning models (ResNet18, ResNet34, ResNet50, EfficientNet_V2_L, Inception_V3, DenseNet121) on six different datasets.
• Evaluation of model performance using F1 scores, recall, and Grad-CAM visualization for duodenal stump (DS) and esophagojejunal (EJ) anastomoses.

Main Results:

• Leakage rates of 1.3% (DS) and 4.3% (EJ) were observed.
• The EXP1 dataset, utilizing single-image analysis with augmentation, yielded the best performance.
• ResNet18 on EXP1 achieved the highest recall (0.8474 for DS, 0.8000 for EJ) and F1 scores (0.6357 for DS, 0.6938 for EJ).
• Grad-CAM analysis indicated the importance of local and surrounding tissue features in predictions.

Conclusions:

• Deep learning models show promise in predicting postoperative leakage during gastric cancer surgery.
• High-resolution imaging, single-image analysis, and data augmentation are key factors for successful model performance.
• These findings support the potential clinical application of AI in surgical image analysis for improved patient care.