人工智能在疟原虫检测中的应用研究

An artificial intelligence model for malaria diagnosis

疟疾是由疟原虫引起的蚊媒传染性疾病,具有分布广泛、传播迅速、潜伏期长等特点,其直接关系到人类的健康、经济的发展以及社会的稳定.疟疾的快速准确检测是降低疟疾的病死率和控制疟疾传播的关键.目前已有研究利用深度学习算法实现对疟原虫的检测,但开发疟疾临床诊断的人工智能系统仍然存在挑战.本研究基于深度学习中的多尺度注意力机制,构建了基于人工智能的疟疾诊断目标检测模型(artificial intelligence-based object detection model for malaria diagnosis,AIM).同时,本研究利用智能手机与光学显微镜收集薄血涂片图像,创建了疟原虫薄血涂片图像数据集(Smart Malaria NET),并用于AIM模型的训练与评估.结果表明,AIM模型的Accuracy为94.49%,Precision为94.54%,Recall为94.49%,F1-score为94.50%,AUC(area under curve)为98.11%,各项评价指标均优于现有的VGG和Res Net模型.该人工智能的疟疾诊断目标检测模型有助于提高缺乏镜检人员地区的疟疾诊断能力,为全球疟疾防控提供“中国技术”与“中国方案”.

Malaria is a mosquito-borne disease caused byPlasmodium,which has a high mortality rate.Rapid and accurate detection ofmalariais the key to reducing malaria mortality and controlling its transmission.Several deep leaming algorithms have previously beenapplied tomalaria blood smearsfordiagnosis usingfeatures extracted frommicroscopic images.Here,we propose a novel artificialintelligence(AI)-based object detection model for malaria diagnosis(AIM)using multi-scale attention approaches.An annotateddataset(SmartMalariaNET)was created consisting of thin smear images acquired by smartphone cameras,which is used to train andassess the AIM.Theresults show that the effectiveness of our model in distinguishing positive and negative images is evidentin thevalues ofthe peromancemetrics,namely Accuracy,Prcision,Rcall,Fcore,and AUC(area under curve),calculated as 94.49%94.54%,94.49%,94.50%,and 98.11%,respectively.AIMis superior to the existing deep leaning modelsin all evaluation indicatorsOur model shows clinically acceptable performance in detecting malaria parasites and could aid in malaria diagnosis in resource-limited regions,especially in areas lacking experienced parasitologists and equipment.