摘要
基于深度网络模型对皮肤病医学图像的精确分割可以为临床诊断提供准确的定位和定量信息,但在训练网络时却需要影像专家对图像进行大量的、专业化的标注。本文结合流形理论与深度学习技术,提出一种弱监督的皮肤病图像分割网络,旨在解决皮肤病图像分割中标注数据稀缺、图像特征复杂多变等问题。该网络通过流形学习捕捉图像数据内在的低维流形结构,有效减少噪声和冗余信息的影响;基于提出的像素相关性模型,分析图像中基层特征之间的空间关系和上下文信息,提升像素级伪标签的准确性;基于注意力机制,提取图像中多种尺度的特征,以适应皮肤病斑块的多样性变化。采用了ISIC-2018皮肤病图像数据集对提出的网络模型进行验证,结果表明,所提出的弱监督皮肤病分割网络在分割精度、鲁棒性和泛化能力等方面均表现出优异的性能。
The accurate segmentation of dermatological medical images based on deep network models can facilitate precise localization and quantitative information for clinical diagnosis.However,training the network necessitates the involvement of image experts who are responsible for labeling images with a substantial number of specialized annotations.This paper presents a novel approach that integrates manifold theory with deep learning technology to develop a weakly supervised dermatological image segmentation network.This methodology addresses the challenges posed by limited labeling data and the intricate and evolving characteristics of dermatological images.The network captures the intrinsic low-dimensional manifold structure of image data through manifold learning,thereby reducing the impact of noise and redundant information.The proposed pixel correlation model is based on the analysis of the spatial relationship and contextual information between the low-level features in the image,with the objective of improving the accuracy of pixel-level pseudo labels.The proposed network model was validated by using the ISIC-2018 skin disease image dataset.The results demonstrated that the network,which was based on the attention mechanism and capable of extracting features of multiple scales in the image,performed well in terms of segmentation accuracy,robustness,and generalization ability.
作者
蔡晟玮
孙玉娟
王桦
张小峰
CAI Shengwei;SUN Yujuan;WANG Hua;ZHANG Xiaofeng(School of Computer and Artificial Intelligence,Ludong University,Yantai 264039,China)
出处
《鲁东大学学报(自然科学版)》
2025年第3期233-242,共10页
Journal of Ludong University:Natural Science Edition
基金
国家自然科学基金(U24A20328,62171209,62272281,61873117)
山东省高校青创团队项目(2023KJ212)。
关键词
图像分割
流形学习
弱监督学习
多尺度可分离注意力
像素相关性
image segmentation
manifold learning
weakly supervised learning
multi-scale separable attention
pixel correlation
