摘要
皮肤病变区域的精准分割是计算机辅助皮肤病诊断的关键性研究任务,它为医生提供了重要的诊断依据。尽管基于UNet的改进网络在图像分割领域取得了显著进展,但是它们往往忽视了掩膜与多层级信息的结合运用,从而限制了分割精度。引入Transformer有助于分割精度的提升,但是网络的参数量也会显著增加,从而可能导致计算复杂度和模型部署成本上升。针对这些问题,提出一种基于多级特征组聚合的皮肤病U型分割网络(MFGA-UNet),其能够充分融合掩膜信息与多层级信息,在保证较低参数量的同时,实现高精度且轻量级的皮肤病图像分割。MFGA-UNet中,首先,采用改进的反瓶颈卷积模块替代标准UNet中的常规卷积块;其次,引入多级特征组聚合模块优化网络的跳跃连接,有效融合掩膜信息与多层级特征,丰富了特征层次;最后,利用深度监督技术改进损失函数,通过解码器各层输出求取损失,优化了模型训练过程。在ISIC2017和ISIC2018皮肤病数据集上的评估结果显示,MFGAUNet的参数量仅有10.246 M,且精度超越了现有6种医学图像分割网络,在两个数据集中,Dice分别达到了94.273%、90.028%。
The precise segmentation of dermatological lesions is a crucial research task in computer-aided skin disease diagnosis,providing vital diagnostic evidence for physicians.Although improved networks based on UNet have made significant progress in the field of image segmentation,they often overlook the integration of mask and multi-level information,thus limiting segmentation accuracy.The introduction of Transformer helps improve segmentation accuracy,but it also significantly increases network parameters,potentially leading to increased computational complexity and model deployment costs.To address these issues,we propose a skin lesion segmentation network based on Multi-level Feature Group Aggregation UNet(MFGA-UNet),which can fully integrate mask information and multi-level information while ensuring a low parameter count,achieving high accuracy and lightweight skin lesion segmentation.In MFGA-UNet,we first replace standard UNet convolution blocks with improved bottleneck convolution modules.Secondly,we introduce a multi-level feature group aggregation module to optimize the network's skip connections,effectively integrating mask information and multi-level features,enriching the feature hierarchy.Finally,we utilize deep supervision techniques to improve the loss function,computing losses from decoder layer outputs,optimizing the model training process.Evaluation on the ISIC2017 and ISIC2018 skin disease datasets shows that MFGA-UNet has only 10.246M parameters and outperforms six existing medical image segmentation networks,with Dice coefficients reaching 94.273% and 90.028% on the two datasets,respectively.
作者
廖骏卿
高琳
邹茂扬
黄海莹
何晋
LIAO Junqing;GAO Lin;ZOU Maoyang;HUANG Haiying;HE Jin(College of Blockchain Technology,Chengdu University of Information Technology,Chengdu 610225,China;The West China Second University Hospital of Sichuan University,Chengdu 610044,China)
出处
《软件导刊》
2025年第8期173-181,共9页
Software Guide
基金
四川省科技计划项目(2020YFS0316)
四川省自然科学基金项目(2023NSFSC0482)。
关键词
图像分割
皮肤病
U型网络
反瓶颈卷积模块
多级特征组聚合
image segmentation
skin lesion
UNet
inverted bottleneck convolution
multi-level feature group aggregation
