改进YOLOv5对病理图像下噪声标签的检测与自动纠正应用

Application of Improved YOLOv5 in the Detection and Automatic Correction of Noise Labels in Pathological Images

病理图像细胞检测是医学诊断的基础部分,正确、精准地检测靶向细胞及其数量对疾病诊疗至关重要。传统医学采用手工镜检的估计方式检测病理图像,依赖病理医生的工作经验,存在主观性、检测精度较低的问题。为此,提出改进YOLOv5的噪声标签检测与自动纠正网络检测病理图像中的靶向细胞,通过Conf、IOU函数使网络具有区分真值标签和噪声标签的能力,从而实现噪声标签的自动纠正,以辅助医生对鼻窦炎疾病类型进行临床诊断。结果表明,改进网络在鼻窦炎病理图像数据集上的平均精度、召回率分别提升至88.9%和95.6%,可满足检测病理图像的精度和纠正噪声标签的需求。

Pathological image cell detection is a fundamental part of medical diagnosis,and accurate detection of targeted cells and their quantities is crucial for disease diagnosis and treatment.Traditional medicine uses manual microscopy to estimate pathological images,relying on the work experience of pathologists,which leads to subjectivity and low detection accuracy.To this end,an improved YOLOv5 noise label detection and automatic correction network is proposed to detect target cells in pathological images.By using Conf and IOU functions,the net⁃work has the ability to distinguish between truth labels and noise labels,thereby achieving automatic correction of noise labels to assist doctors in clinical diagnosis of sinusitis disease types.The results showed that the improved network achieved an average accuracy and recall rate of 88.9%and 95.6%respectively on the pathological image dataset of sinusitis,which can meet the requirements of detecting pathological images and correcting noise labels.