摘要
新生儿胆道闭锁是新生儿常见的致命疾病之一,并且该病在亚洲的发病率高于世界其他地区.新生儿胆道闭锁需要及时发现及时治疗,然而由于缺少专业的儿科医生和辅助诊疗手段,新生儿父母往往不能及时发现而错过了最佳治疗时间.因此,本文开发了一个具有实际应用价值的预诊算法,通过新生儿粪便图片预测新生儿是否患有新生儿胆道闭锁,并提醒新生儿父母及时就诊.为了让算法在应用场景下识别率更高,本文的算法基于一个真实场景下拍摄的新生儿粪图片数据集开发.首先我们设计了一个自注意力网络模型BANet(Biliary Atresia Network),将图片的浅层特征和深层特征相结合,可以得到更好的分类效果.由于拍摄自应用场景下的图片存在过暗和过曝等问题.通过分析数据集的亮度分布,我们设计了一个自动亮度调节算法解决.此外,图片中的阴影也会对识别结果造成干扰,因此我们在训练阶段增加了一种阴影数据增强方式来缓解这一问题.为验证本文提出算法的有效性,本文设计了一个和医生的对比试验.结果证明BANet在四分类的识别率、二分类的识别率、特异性和敏感性等客观评价指标上占有明显优势.本文提出的BANet能够有效利用图片中的颜色、异常点等信息,并且通过对图片亮度进行补偿,提高了整体算法准确率和鲁棒性,在实际应用场景取得了良好的效果.
Neonatal biliary atresia is one of the most common fatal diseases in neonates,with higher incidence rates in Asia than in other parts of the world.Early detection and treatment of neonatal biliary atresia are crucial,yet the lack of professional pediatricians and auxiliary diagnostic and treatment methods can cause parents to miss the best treatment window.To address this issue,this paper develops a predictive algorithm with practical application value that uses neonatal stool pictures to predict whether the newborn has neonatal biliary atresia and reminds parents to visit a doctor in time.To achieve higher recognition rates in practical scenarios,the algorithm in this paper is developed based on a real-scene dataset of newborn fecal images.First,we designed a self-attention network model BANet(Biliary Atresia Network),which will combine shallow features and deep features of pictures to get better classification.To address issues as dark or overexposed images,we developed an automatic brightness adjustment algorithm by analyzing the brightness distribution of the dataset.Furthermore,we added a shadow data enhancement method duiring training to mitigate the inference of shadows on recognition results.In order to verify the effectiveness of the algorithm proposed in this paper,we design a comparison test with doctors.The results show that BANetoutperformed doctors in objective evaluation indicators such as the recognition rate of four classifications,the recognition rate of two classifications,specificity and sensitivity.The proposed BANet can effectively use the color,abnormal points and other information in the picture,by compensating the brightness of the picture,the accuracy and robustness of the overall algorithm are improved and good results have been achieved in practical application scenarios.
作者
秦中翰
艾成博
谭超群
刘洪
杜文超
杨红雨
吴志红
陈虎
QIN Zhong-Han;AI Cheng-Bo;TAN Chao-Qun;LIU Hong;DU Wen-Chao;YANG Hong-Yu;WU Zhi-Hong;CHEN Hu(College of Computer Science,Sichuan University,Chengdu 610065,China;West China Hospital of Sichuan University,Chengdu 610065,China;State Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610065,China)
出处
《四川大学学报（自然科学版）》
CAS
CSCD
北大核心
2023年第6期72-81,共10页
Journal of Sichuan University(Natural Science Edition)
基金
国家自然科学基金(61871277)
四川省卫生健康委员会科研课题(19PJ007)
成都市卫生健康委员会科研课题(2022053)。
关键词
新生儿胆道闭锁
自注意力
数据增强
卷积神经网络
Neonatal biliary atresia
Self-attention
Data augmentation
Convolutional neural network
