A Deep Learning Model Incorporating U-Net Image Segmentation And Yolov5 Algorithm To Assess Preoperative Surgical Handwashing Specification In Ophthalmology

Based on the operating room surveillance video, a deep learning model integrating U-Net image segmentation and YOLOv5 algorithm was constructed to evaluate the accuracy and standardization of preoperative surgical handwashing action recognition by ophthalmology nurses.

artificial intelligence 

2,200 copies of preoperative surgical handwashing surveillance videos in the handwashing area of the operating room of Wuhan Aier Eye Hospital were retrieved, among which 200 copies of standardized handwashing/hand brushing videos were standardized and categorized for the standardized action, which was disassembled into 25 action classifications and 5 item classifications. Meanwhile, the other 2000 video materials are divided into training set, validation set and test set according to the ratio of 8:1:1. The YOLOv5 target detection algorithm and U-Net image segmentation technique were integrated to construct the deep learning model. 

 Action and item recognition accuracy analysis results show that the model's mean average precision (mAP) in single-person scenarios performs near-optimally under different intersection and merger ratio thresholds, with mAP reaching 0.9810 when the threshold is 0.5. In addition, the mean F1, precision, and recall values for all categorized item recognitions are 0.9800, 1.0000, and 0.9900, respectively. The mAP value of the deep learning model for action detection in multi-person scenario reaches 0.8699 at 0.5 threshold intersection ratio, and the average F1 value (0.8760), precision value (0.8937) and recall value (0.8699) are lower than that of the detection results in the single-player scenario.

 Based on the operating room surveillance video, the deep learning model integrating U-Net image segmentation and YOLOv5 algorithm can effectively monitor the standardization of preoperative surgical handwashing by ophthalmology nurses with high accuracy.