摘要
目前,口腔疾病的患病率和发病率逐年增长,口腔疾病患病者术后缺乏定期检查意识导致病情复发,早期发现可以提前预防口腔疾病的发作或者降低治疗难度,术后定期检查可以降低病情复发概率。然而,口腔疾病病例特征复杂,仅靠口腔医师人工阅片存在个人主观误差,导致误诊漏诊。尽管各医学领域的自动诊断取得了进步,但由于相关数据集的缺乏,加上口腔异常的多样性和复杂性,口腔异常目标检测仍是一项巨大挑战。针对当前口腔相关数据集较少的问题,提出了一个新的口腔全景X光片数据集,其中包含1600张具有标准诊断注释标签的口腔全景X光片。此外,还针对该数据集提出了一个深度学习目标检测框架以实现9类口腔疾病智能化、自动化诊断。ResNet提取口腔疾病特征,双向注意力特征金字塔融合网络从多层次特征图中自适应捕获丰富的上下文信息,从而提高不同尺度的口腔异常疾病检测泛化能力,RPN网络进行口腔疾病区域生成,回归分类网络中通过一个多标签损失函数以降低标签竞争。该框架在所提出的数据集上可以准确地检测9类口腔疾病,各类疾病检测平均精度较基础模型都有不同程度的提升,mAP达82.5%。
Current research indicates a year-over-year increase in the prevalence and incidence of oral diseases.Additionally,postoperative patients with oral diseases often lack awareness of regular check-ups,leading to disease recurrence.Early detection can prevent the onset of oral diseases or reduce the difficulty of treatment.Regular postoperative examinations can lower the probability of disease recurrence.However,oral disease cases are characterized by complexity,and relying solely on manual radiograph review by oral healthcare professionals introduces subjective errors,leading to misdiagnosis or missed diagnoses.Despite progress in automated diagnostics in medicine,limited datasets and the complex nature of oral abnormalities pose significant challenges for oral anomaly detection.Addressing the limited availability of oral-related datasets,this paper proposes a new panoramic dental X-ray dataset comprising 1600 images,each annotated with standard diagnostic labels.The paper introduces a deep learning object detection framework for intelligent and automated diagnosis of nine types of oral diseases.The framework utilizes ResNet for extracting oral disease features,a bidirectional attention feature pyramid fusion network to adaptively capture rich contextual information from multi-level feature maps,enhancing the generalization capability for detecting oral abnormalities of varying scales.The Region Proposal Network(RPN)generates regions of oral disease,and in the regression classification network,a multi-label loss function is proposed to mitigate label competition.The proposed framework achieves accurate detection of the nine oral diseases on the provided dataset,with varying degrees of improvement in mean Average Precision(mAP)compared to baseline models,The mAP reaches 82.5%.
作者
余红兵
扶思思
陈嘉成
赵梦婷
黄武秋
YU Hongbing;FU Sisi;CHEN Jiacheng;ZHAO Mengting;HUANG Wuqiu(Shenzhen Nanshan Center for Chronic Disease Control,Shenzhen 518054,China;School of Software,South China Normal University,Foshan 528225,China;School of Computer Science and Technology,Harbin Institute of Technology(Shenzhen),Shenzhen 518055,China)
出处
《软件导刊》
2025年第9期174-180,共7页
Software Guide
基金
国家自然科学基金项目青年科学基金项目(62302172)。
