一种基于深度学习的新型针灸针自动计数的云服务系统:提高针灸安全性的策略  

A novel deep learning based cloud service system for automated acupuncture needle counting:a strategy to improve acupuncture safety

作　　者：黄梓皓 韦君逸 陈海勇[1] 吴凤亮 WONG Tsz Ho;WEI Junyi;CHEN Haiyong;NG Bacon Fung Leung(School of Chinese Medicine,Li Ka Shing Faculty of Medicine,University of Hong Kong,Hong Kong 999077,China;Department of Mathematics and Statistics,Georgetown University,Washington D.C.20057,USA;Department of Rehabilitation Sciences,The Hong Kong Polytechnic University,Hong Kong 999077,China)

机构地区：[1]香港大学李嘉诚医学院中医药学院,中国香港999077 [2]Department of Mathematics and Statistics,Georgetown University,Washington D.C.20057,USA [3]香港理工大学康复科学系,中国香港999077

出　　处：《Digital Chinese Medicine》2024年第1期40-46,共7页数字中医药（英文）

摘　　要：目的病人针灸治疗时,漏针可能造成严重后果。为增强针灸安全性,本研究将开发一个基于深度学习的自动云端数针系统。方法本研究利用拍有针灸针的手机图片为曾预先训练的定向目标检测模型(Oriented R-CNN)进行转移学习,以开发一种能够自动计算图片中针灸针数量的目标检测算法。我们首先拍摄了一个包含590张图片的训练集以及一个包含1025张图片的验证集。随后在带有NVIDIA Tesla T4图形处理器(GPU)的Google Colab环境中部署了MMRotate工具箱为模型进行训练。我们将训练的模型与Telegram bot手机小程序界面集成,并评估模型的准确率、精确率和召回率。本小程序中云端数针系统的速度则以端到端处理时间计算。结果在每张图片20根针的验证集情景中,我们的定向检测模型在准确率、精确率和召回率上分别达到96.49%、99.98%和99.84%,平均端到端处理时间为1.535秒。结论这项云服务系统在速度、准确性和可靠性上的提升展示了深度学习物体检测技术改进针灸实践的潜力。Objective The unintentional retention of needles in patients can lead to severe conse-quences.To enhance acupuncture safety,the study aimed to develop a deep learning-based cloud system for automated process of counting acupuncture needles.Methods This project adopted transfer learning from a pre-trained Oriented Region-based Convolutional Neural Network(Oriented R-CNN)model to develop a detection algorithm that can automatically count the number of acupuncture needles in a camera picture.A train-ing set with 590 pictures and a validation set with 1025 pictures were accumulated for fine-tuning.Then,we deployed the MMRotate toolbox in a Google Colab environment with a NVIDIA Tesla T4 Graphics processing unit(GPU)to carry out the training task.Furthermore,we integrated the model with a newly-developed Telegram bot interface to determine the ac-curacy,precision,and recall of the needling counting system.The end-to-end inference time was also recorded to determine the speed of our cloud service system.Results In a 20-needle scenario,our Oriented R-CNN detection model has achieved an accu-racy of 96.49%,precision of 99.98%,and recall of 99.84%,with an average end-to-end infer-ence time of 1.535 s.Conclusion The speed,accuracy,and reliability advancements of this cloud service system innovation have demonstrated its potential of using object detection technique to improve acupuncture practice based on deep learning.

关 键 词：人工智能 电脑视觉 目标检测 针灸 病人安全 

分 类 号：R245[医药卫生—针灸推拿学] TP18[医药卫生—中医临床基础]