摘要
由于皮肤病的类内差异大、类间差异小及样本分布不均衡导致恶性皮肤病智能诊断误诊率极高,因此提出一种基于深度残差金字塔的皮肤病变分割与特征提取机制。具体地,对比主流的单一尺度网络输出,为了提高特征提取和训练结果的准确度,构建了一个深度残差金字塔多尺度编码网络,将瓶颈层划分为多尺度编码网络,通过提取多尺度特征,实现网络分割与提取结果的输出。进一步,为了解决样本不均衡问题,设计了基于焦点损失的损失梯度监督机制,即通过焦点损失增加模型对难分样本的关注度,同时通过梯度协调机制减小难分样本和离群点对模型整体准确率的影响,从而达到减小类别不平衡对诊断结果的影响。实验结果表明,所提机制的分割与提取与现有相关方案相比,Jaccard系数提高了3%~10%,达到82.3%。
The lesion appearance of skin diseases is characterized by small inter-class variance,large intra-class variance and imbalanced sample distribution,which results in a high misdiagnosis rate of intelligent diagnosis.A deep residual pyramid-based skin lesion segmentation and feature extraction mechanism is proposed to address the above problem.Specifically,compared with the main-stream single-scale network output,a deep residual pyramid multi-scale coding network is constructed to increase the accuracy of feature extraction and training outcomes,and the bottleneck layer is divided into multi-scale coding networks to segment and extract multi-scale features.Meanwhile,a loss gradient supervision mechanism based on focal loss is designed to solve the problem of sample imbalance.That is,the focal loss is used to increase the model's attention to the hard-to-separate samples,and the gradient coordination mechanism is developed to reduce the impact of hard-to-separate samples and outliers,so as to reduce the impact of sample imbalance on the diag-nosis results.Finally,the experiment results demonstrate that the proposed mechanism of segmentation and extraction obtains competitive accuracy compared with existing representative works and the Jaccard index is 3%-10%higher,reaching 82.3%.
作者
董春序
李雪
陈思光
DONG Chunxu;LI Xue;CHEN Siguang(School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing Jiangsu 210003;Department of Dermatology,Womens Hospital of Nanjing Medical University(Nanjing Woman and Childrens healthcare Hospital),Nanjing Jiangsu 210004)
出处
《传感技术学报》
2025年第2期263-271,共9页
Chinese Journal of Sensors and Actuators
基金
国家自然科学基金项目(61971235)
中国博士后科学基金(面上一等资助)项目(2018M630590)
江苏省“333高层次人才培养工程”项目
江苏省博士后科研资助计划项目(2021K501C)
南京邮电大学‘1311’人才计划和赛尔网络下一代互联网技术创新项目(NGII20190702)。
