基于深度学习技术和计算机语言设计烧烫伤创面深度智能检测模型并予以测试，验证该模型对烧烫伤创面图像检测的有效性及准确性。

收集2022年1月至2024年2月在上海交通大学医学院附属瑞金医院烧伤整形科接受治疗且符合入选标准的烧烫伤患者伤后48 h内创面照片共492张，重置照片顺序并编号。由两名执业3年以上的副主任医师采用图像标注工具LabelMe对照片中的目标创面进行标记并判定其严重程度，严重程度分为Ⅰ度、Ⅱ度和Ⅲ度。采用图像处理技术扩充数据集至2 952张，按照7∶2∶1比例划分为训练集、验证集及测试集。在Python 3.10.0版本下，提出并构建基于深度学习的烧烫伤创面深度智能检测模型P-YOLO，通过多批次训练调整并优化网络参数。经过测试得到该模型在数据集上的各项指标参数，如查准率、召回率以及在不同交并比(IoU)下的平均精度均值等，根据实验结果绘制出相应的F1指数曲线和混淆矩阵。

(1)经测试，所设计的P-YOLO智能检测模型对Ⅰ度、Ⅱ度和Ⅲ度烧烫伤创面识别的查准率分别为0.962、0.931及0.886，召回率分别为0.849、0.828及0.857，F1指数分别为0.902、0.876及0.871。(2)混淆矩阵显示，P-YOLO模型检测Ⅰ度、Ⅱ度和Ⅲ度烧烫伤创面的准确度分别为0.86、0.87及0.91。(3)当IoU阈值为0.5时，P-YOLO模型检测Ⅰ度、Ⅱ度和Ⅲ度烧烫伤创面的平均精度均值为0.893、0.885及0.838。在所有类别创面中，P-YOLO模型检测的平均精度均值为0.872。(4)与Faster R-CNN、YOLOv5及YOLOv7检测模型相比，P-YOLO模型具有最高的平均精度均值，检测效果最优。

基于深度学习的智能检测模型P-YOLO整体检测准确率和可靠性较高，能够提高烧伤科医师对烧烫伤创面深度的诊断准确度和效率。

To develop an intelligent detection model for burn wound depth based on deep learning technology and computer vision, and test its efficiency and accuracy in detecting burn wound depth.

From January 2022 to February 2024, 492 burn wound photos meeting the inclusion criteria were collected from patients treated at the Department of Burn and Plastic Surgery in Ruijin Hospital, School of Medicine, Shanghai Jiao Tong University. The photos were reordered and numbered. Two associate physicians who had been practing for more than three yeas in the Burn and Plastic Surgery Department used the LabelMe image annotation tool to mark the target wounds and determine the severity, categorized into first-degree, second-degree, and third-degree. The dataset was expanded to 2 952 images using image processing techniques and was split into training, validation, and test sets in a 7∶2∶1 ratio. An intelligent burn detection model P-YOLO based on deep learning was proposed and constructed under Python 3.10.0, with network parameters adjusted and optimized through multiple batches of training. The model's performance was tested, yielding various metrics such as precision, recall, and mean average precision (mAP) under different intersection over union (IoU) thresholds. Corresponding F1 score curves and confusion matrices were plotted based on these experimental results.

(1)Testing results showed that the precision rates of the designed P-YOLO intelligent detection model for identifying first-degree, second-degree, and third-degree burns were 0.962, 0.931, and 0.886, respectively. The recall rates were 0.849, 0.828, and 0.857, respectively, and the F1 scores were 0.902, 0.876, and 0.871, respectively. (2)The confusion matrix indicated that the P-YOLO model correctly detected first-degree, second-degree, and third-degree burns with accuracy of 0.86, 0.87 and 0.91, respectively. (3)When the IoU threshold was 0.5, the mAP of P-YOLO model for detecting first-degree, second-degree, and third-degree burns was 0.893, 0.885 and 0.838, respectively. The model achieved an overall mAP of 0.872 across all categories. (4)Compared to Faster R-CNN, YOLOv5, and YOLOv7 detection models, P-YOLO model had the highest mAP and exhibited the best detection performance.

The deep learning-based intelligent detection model P-YOLO shows high overall detection accuracy and reliability, which can significantly enhance the accuracy and efficiency of burn wound depth diagnosis for burn surgeons.