基于YOLO神经网络构建压力性损伤自动检测和分期的人工智能模型

Construction of an Artificial Intelligence-assisted System for Automatic Detection of Pressure Injury Based on the YOLO Neural Network

背景随着人口老龄化,压力性损伤(PI)的发病率逐渐增加,这不仅严重影响了患者的生存质量,还增加了医保支出。然而,PI的早期发现和准确分期极大地依赖于专业培训。目的构建并测试一个用于PI自动检测和分期的人工智能模型,以提高PI诊断的实时性、准确性和客观性。方法选取常熟市第一人民医院压疮电子化管理系统中2021年1月—2024年2月的693张PI图像,将图像随机划分为训练集(551张)和测试集(142张),并按照2019年美国压疮咨询委员会(NPUAP)制订的PI预防和治疗指南分为6期,包括:Ⅰ期154张、Ⅱ期188张、Ⅲ期160张、Ⅳ期82张、深部组织损伤期57张、不可分期52张。利用基于5种不同版本的YOLOv8[nano(n)、small(s)、medium(m)、large(l)和extra large(x)]神经网络和迁移学习,建立针对PI的深度学习目标检测模型。模型评价指标包括精确度、准确率、灵敏度、特异度及检测速度等。最后,通过Ultralytics Hub平台将模型部署到手机应用程序(App)中,实现AI模型在临床工作中的应用。结果在对包含142张PI图像的测试集进行评估时,YOLOv8l版本在确保高精确度(0.827)的同时,也展现了较快的推理速度(68.49帧/s),与其他YOLO版本相比,在精确度与速度之间取得了最佳的平衡。具体而言,其在所有类别上的整体准确率为93.18%,灵敏度为76.52%,特异度为96.29%,假阳性率为3.72%。在6个PI分期中,模型预测Ⅰ期的准确率最高,达到95.97%;预测Ⅱ期、Ⅲ期、Ⅳ期、深部组织损伤期、不可分期分别取得了91.28%、91.28%、91.95%、95.30%和93.29%的准确率。就处理速度而言,YOLOv8l处理142张图像的总耗时为2.07 s,平均每秒可处理68.49张PI图像。结论基于YOLOv8l网络的AI模型能够快速、准确地对PI进行检测和分期。将该模型部署到手机App中,能够在临床实践中便携使用,具有很大的临床应用潜力。

Background With the aging population,the incidence of pressure injury(PI)is gradually increasing.This not only severely impacts the quality of life for patients but also increases healthcare expenditures.However,the early detection and accurate staging of PI heavily depend on specialized training.Objective To construct and validate an artificial intelligence model for the automatic detection and staging of PI aimed at enhancing the real-time nature,accuracy,and objectivity of PI diagnostics.Methods A total of 693 PI images from the electronic management system of pressure ulcers at Changshu No.1 People's Hospital were selected from January 2021 to February 2024,the images were randomly divided into a training set(551 images)and a test set(142 images),and categorized into six stages according to National Pressure Ulcer Advisory Panel(NPUAP)guidelines:StageⅠ(154 images),StageⅡ(188 images),StageⅢ(160 images),StageⅣ(82 images),deep tissue injury(57 images),and unstageable(52 images).A deep learning object detection model for PI was established using five different versions of the YOLOv8[nano(n),small(s),medium(m),large(l)and extra large(x)]neural network and transfer learning.The model evaluation metrics included accuracy,sensitivity,specificity,false positive rate,and detection speed.Finally,the model was deployed to a mobile application via the Ultralytics Hub platform,facilitating the application of the AI model in clinical practice.Results During the evaluation of a test set containing 142 PI images,the YOLOv8l version demonstrated high accuracy(0.827)and fast inference speed(68.49 fps),achieving the best balance between precision and speed among the YOLO versions.Specifically,it achieved an overall accuracy of 93.18%across all categories,a sensitivity of 76.52%,a specificity of 96.29%,and a false positive rate of 3.72%.Among the six stages of PI,the model achieved the highest accuracy for StageⅠat 95.97%.The accuracies for StageⅡ,StageⅢ,StageⅣ,deep tissue injury,and unstageable were 91.28%,91.28%,91.95%,95.30%,and 93.29%,respectively.In terms of processing speed,YOLOv8l took a total of 2.07 seconds to process 142 images,averaging 68.49 PI images per second.Conclusion The AI model based on the YOLOv8l network can quickly and accurately detect and stage PI.Deploying this model to a mobile app allows for portable use in clinical practice,demonstrating significant potential for clinical application.