Dense-UNet:基于密集连接U-Net的创伤伤口图像分割  

Dense-UNet:Trauma image segmentation based on densely connected U-Net

作　　者：张涛 禹宝庆 雷鸣 高洁 胡伏原[2] 潘云峰 周涛 ZHANG Tao;YU Baoqing;ILEI Ming;GAO Jie;HU Fuyuan;PAN Yunfeng;ZHOU Tao(Shanghai Seventh People's Hospital,Shanghai 200120,China;School of Electronic&Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China;School of Computer Science&Engineering,North Minzu University,Yinchuan 750004,China)

机构地区：[1]上海第七人民医院,上海200120 [2]苏州科技大学电子与信息工程学院,苏州215009 [3]北方民族大学计算机科学与工程学院,银川750004

出　　处：《中国体视学与图像分析》2024年第4期282-289,共8页Chinese Journal of Stereology and Image Analysis

基　　金：国家自然科学基金项目(No.82174189,82304920);上海市浦东新区卫生健康委员会学科建设(PWxq2022-09);上海市浦东新区卫生健康委员会卫生计生科研项目(PW2021D-05);上海市第七人民医院医工交叉创新专项项目QYYGJO111。

摘　　要：慢性伤口会严重的影响患者生活质量,并且病情容易恶化。为保证伤口图像分割的分割效果,提高基于图像的伤口分析能力,本文提出了一种基于密集连接的U-Net模型用于创伤伤口分割。其中,在跳跃连接中应用密集连接机制,将不同阶段的编码器输出,通过密集连接机制汇聚至对应层的解码器中。在每一层的解码器中使用多视野特征自适应融合模块(MFAF),对每一层解码器的跳跃连接特征进行自适应融合。在临床公共数据集上验证了本文方法的有效性。对比实验结果表明,本文方法对创伤伤口分割的DSC、MIoU、HD95、Recall、Voe及Rvd结果分别为82.84%、74.06%、2.66%、85.10%、74.31%和73.12%。并且对于伤口边缘不清晰的情况,分割精度得到有效提升。Chronic wounds not only seriously affect the quality of life of patients,but also prone to deterioration.To ensure the accuracy of wound image segmentation and improve the image-based wound analysis,a U-Net model based on dense connections is proposed for wound segmentation in this paper.The dense connection mechanism employs skip connections.The outputs of different stages in the encoder are aggregated to the decoder by the dense connection mechanism.Within each layer of the decoder,the Multi-view Feature Adaptive Fusion module(MFAF)is utilized to adaptively fuse the skip connection features from each layer.The effectiveness of the proposed method is validated on public clinical datasets.Comparative experimental results show that the DSC,MIoU,HD95,Recall,Voe and Rvd metrics of the proposed method for wound segmentation are 82.84%,74.06%,2.66%,85.10%,74.31% and 73.12%,respectively.Moreover,the segmentation accuracy is effectively improved for the cases with unclear wound edges.

关 键 词：U-Net 创伤伤口分割 密集连接 自适应特征融合 多尺度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] R473.6[医药卫生—护理学]