一种对噪声鲁棒的皮肤病变分割网络  

A noise-robust skin lesion segmentation network

作　　者：程文豪 朱戈[1] 唐锦萍 CHENG Wenhao;ZHU Ge;TANG Jinping(School of Computer and Big Data,Heilongjiang University,Harbin 150080,China)

机构地区：[1]黑龙江大学计算机与大数据学院,哈尔滨150080

出　　处：《黑龙江大学自然科学学报》2025年第1期116-126,共11页Journal of Natural Science of Heilongjiang University

基　　金：黑龙江省重点研发计划指导类项目(GZ20230015);黑龙江省自然科学基金资助项目(LH2022F044);黑龙江省高等教育本科教育教学改革研究重点委托项目(SJGZ20220050);黑龙江省省属高等学校基本科研业务费项目(2022-KYYWF-1047,2023-KYYWF-1467)。

摘　　要：为了解决噪声标签问题,现有的医学图像分割方法常从数据集中分离出正确标记的图像和噪声标签对应的图像,使用正确标签训练网络,采用噪声标签给网络提供额外的信息。此方法,获取绝对正确的标签是十分困难的。因此,将数据集中的所有标签都视作带噪声的标签,利用框架中均值教师(Mean Teacher,MT)网络比学生网络更稳定的特点,引导学生网络抵消标签中固有噪声对预测结果的影响,以此实现对噪声的鲁棒。同时,为了进一步提升分割性能,提出了一种对噪声鲁棒的皮肤病变分割网络(Noise-robust skin lesion segmentation network,NSLS-Net),它主要由多尺度分组聚合模块(Multiscale group aggregation Module,MSGAM)和对噪声鲁棒的边缘损失模块(Boundary loss module,BLM)两部分组成。多尺度分组聚合模块采用分组聚合的方式融合了不同尺度下的特征图,以此准确分割具有多样性的皮肤病。边缘损失模块采用加权的方式结合平均绝对误差(Mean absolute error,MAE)和联合差异(Difference over union,DoU)损失,使网络精确分割出病变的边缘区域。在ISIC 2016、PH2、ISIC 2017和ISIC 2018四个皮肤病变数据集上的实验结果表明,所提出的NSLS-Net优于大部分先进的方法,特别是在PH2数据集上,其骰子系数(Dice coefficent,Dice)和交并比(Intersection over union,IOU)指标分别达到了92.80%和87.20%。To address the issue of noisy labels,existing medical image segmentation methods typically first separate correctly labeled images from those with noisy labels in the datasets.The network is then trained with the correct labels,while the noisy labels provide additional information.However,It is difficult to obtain absolutely correct labels.Therefore,this paper treats all labels in the datasets as noisy labels and leverages the stability of the teacher network in the Mean Teacher(MT)framework to guide the student network in mitigating the inherent noise in the labels,thereby achieving robustness against noise.Furthermore,to enhance segmentation performance,the Noise-robust skin lesion segmentation network(NSLS-Net),is introduced which mainly consists of the Multi-scale group aggregation module(MSGAM)and the Noise-robust boundary loss module(BLM).MSGAM aggregates feature maps from different scales using a grouped approach to accurately segment diverse skin lesions.BLM combines Mean absolute error(MAE)with Difference over union loss(DoU)in a weighted manner,enabling the network to precisely segment the boundary regions of lesions.The experimental results on four skin lesion datasets,ISIC 2016,PH2,ISIC 2017,and ISIC 2018,show that the proposed NSLS-Net outperforms the state-of-the-art methods,especially on the PH2 dataset,where the Dice coefficient(Dice)and Intersection over union(IOU)metrics reach 92.80%and 87.20%,respectively.

关 键 词：医学图像分割 皮肤病变 噪声标签 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]