An aerial photographic coverage acquired on two consecutive days in October 2021 with a ground resolution of 20 cm and a spectral resolution of 4 bands (red, green, blue and near infrared), allowed to distinguish most...An aerial photographic coverage acquired on two consecutive days in October 2021 with a ground resolution of 20 cm and a spectral resolution of 4 bands (red, green, blue and near infrared), allowed to distinguish most of the classes of interest present in the intertidal zone of the Sado estuary. We explored the possibilities of thematic classification in the powerful and complex software ArcGIS Pro;we presented the methodology used in a detailed way that allows others with minimal knowledge of GIS to reproduce the classification process without having to decipher the specifics of the software. The classification implemented used ground truth from four classes related to the macro-occupations of the area. In a first phase we explore the standard algorithms with object-based capabilities, like K-Nearest Neighbor, Random Trees Forest and Support Vector Machine, and in a second phase we proceed to test three deep learning classifiers that provide semantic segmentation: a U-Net configuration, a Pyramid Scene Parsing Network and DeepLabV3. The resulting classifications were quantitatively evaluated with a set of 500 control points in a test area of 37,500 × 12,500 pixels, using confusion matrices and resorting to Cohen’s kappa statistic and the concept of global accuracy, achieving a Kappa in the range [0.72, 0.81] and a global accuracy between 88.9% and 92.9%;the option U-Net had the most interesting results. This work establishes a methodology to provide a baseline for assessing future changes in the distribution of Sado estuarine habitats, which can be replicated in other wetland ecosystems for conservation and management purposes.展开更多
针对由于血管类间具有强相似性造成的动静脉错误分类问题,提出了一种新的融合上下文信息的多尺度视网膜动静脉分类网络(multi-scale retinal artery and vein classification network,MCFNet),该网络使用多尺度特征(multi-scale feature...针对由于血管类间具有强相似性造成的动静脉错误分类问题,提出了一种新的融合上下文信息的多尺度视网膜动静脉分类网络(multi-scale retinal artery and vein classification network,MCFNet),该网络使用多尺度特征(multi-scale feature,MSF)提取模块及高效的全局上下文信息融合(efficient global contextual information aggregation,EGCA)模块结合U型分割网络进行动静脉分类,抑制了倾向于背景的特征并增强了血管的边缘、交点和末端特征,解决了段内动静脉错误分类问题。此外,在U型网络的解码器部分加入3层深度监督,使浅层信息得到充分训练,避免梯度消失,优化训练过程。在2个公开的眼底图像数据集(DRIVE-AV,LES-AV)上,与3种现有网络进行方法对比,该模型的F1评分分别提高了2.86、1.92、0.81个百分点,灵敏度分别提高了4.27、2.43、1.21个百分点,结果表明所提出的模型能够很好地解决动静脉分类错误的问题。展开更多
目的甲状腺结节的精准分割在医学影像处理中具有重要意义,然而,超声图像中的结节通常具有尺寸多变和边缘模糊的特点,这为其准确分割带来了挑战。为有效应对这一挑战,本文提出了一种结合卷积神经网络(convolutional neural network,CNN)...目的甲状腺结节的精准分割在医学影像处理中具有重要意义,然而,超声图像中的结节通常具有尺寸多变和边缘模糊的特点,这为其准确分割带来了挑战。为有效应对这一挑战,本文提出了一种结合卷积神经网络(convolutional neural network,CNN)和Transformer的分割网络,命名为TransUNet,旨在实现对甲状腺结节超声图像的精准分割。方法首先,使用卷积神经网络对超声图像进行编码,以生成特征图。接着,将特征图转换为序列向量,并利用Transformer的编码机制来捕捉上下文信息。此外,为保持局部细节特征的完整性,研究组还引入了跳跃连接,将其用于在解码器中对编码特征进行上采样,这对于处理边缘模糊等问题尤为重要。结果通过在甲状腺结节图像分割任务中进行广泛的实验,验证TransUNet的有效性。具体而言,骰子系数(dice coefficient,DICE)为0.75,交并比(intersection over union,IoU)为0.60,F1分数(F1 Score)为0.72,准确率高达0.93,AUC(area under the ROC curve)为0.91。这些性能指标反映了该方法在处理尺寸多变和边缘模糊等挑战方面的出色表现。结论本文提出的TransUNet为甲状腺结节超声图像分割任务带来了显著的性能提升。相较于传统的U-Net方法,TransUNet不仅更好地处理了尺寸多变和边缘模糊等挑战,而且在分割性能上具有更为出色的表现,为医学图像处理领域的进一步研究和临床应用提供了有力支持。展开更多
文摘An aerial photographic coverage acquired on two consecutive days in October 2021 with a ground resolution of 20 cm and a spectral resolution of 4 bands (red, green, blue and near infrared), allowed to distinguish most of the classes of interest present in the intertidal zone of the Sado estuary. We explored the possibilities of thematic classification in the powerful and complex software ArcGIS Pro;we presented the methodology used in a detailed way that allows others with minimal knowledge of GIS to reproduce the classification process without having to decipher the specifics of the software. The classification implemented used ground truth from four classes related to the macro-occupations of the area. In a first phase we explore the standard algorithms with object-based capabilities, like K-Nearest Neighbor, Random Trees Forest and Support Vector Machine, and in a second phase we proceed to test three deep learning classifiers that provide semantic segmentation: a U-Net configuration, a Pyramid Scene Parsing Network and DeepLabV3. The resulting classifications were quantitatively evaluated with a set of 500 control points in a test area of 37,500 × 12,500 pixels, using confusion matrices and resorting to Cohen’s kappa statistic and the concept of global accuracy, achieving a Kappa in the range [0.72, 0.81] and a global accuracy between 88.9% and 92.9%;the option U-Net had the most interesting results. This work establishes a methodology to provide a baseline for assessing future changes in the distribution of Sado estuarine habitats, which can be replicated in other wetland ecosystems for conservation and management purposes.
文摘针对由于血管类间具有强相似性造成的动静脉错误分类问题,提出了一种新的融合上下文信息的多尺度视网膜动静脉分类网络(multi-scale retinal artery and vein classification network,MCFNet),该网络使用多尺度特征(multi-scale feature,MSF)提取模块及高效的全局上下文信息融合(efficient global contextual information aggregation,EGCA)模块结合U型分割网络进行动静脉分类,抑制了倾向于背景的特征并增强了血管的边缘、交点和末端特征,解决了段内动静脉错误分类问题。此外,在U型网络的解码器部分加入3层深度监督,使浅层信息得到充分训练,避免梯度消失,优化训练过程。在2个公开的眼底图像数据集(DRIVE-AV,LES-AV)上,与3种现有网络进行方法对比,该模型的F1评分分别提高了2.86、1.92、0.81个百分点,灵敏度分别提高了4.27、2.43、1.21个百分点,结果表明所提出的模型能够很好地解决动静脉分类错误的问题。
文摘目的甲状腺结节的精准分割在医学影像处理中具有重要意义,然而,超声图像中的结节通常具有尺寸多变和边缘模糊的特点,这为其准确分割带来了挑战。为有效应对这一挑战,本文提出了一种结合卷积神经网络(convolutional neural network,CNN)和Transformer的分割网络,命名为TransUNet,旨在实现对甲状腺结节超声图像的精准分割。方法首先,使用卷积神经网络对超声图像进行编码,以生成特征图。接着,将特征图转换为序列向量,并利用Transformer的编码机制来捕捉上下文信息。此外,为保持局部细节特征的完整性,研究组还引入了跳跃连接,将其用于在解码器中对编码特征进行上采样,这对于处理边缘模糊等问题尤为重要。结果通过在甲状腺结节图像分割任务中进行广泛的实验,验证TransUNet的有效性。具体而言,骰子系数(dice coefficient,DICE)为0.75,交并比(intersection over union,IoU)为0.60,F1分数(F1 Score)为0.72,准确率高达0.93,AUC(area under the ROC curve)为0.91。这些性能指标反映了该方法在处理尺寸多变和边缘模糊等挑战方面的出色表现。结论本文提出的TransUNet为甲状腺结节超声图像分割任务带来了显著的性能提升。相较于传统的U-Net方法,TransUNet不仅更好地处理了尺寸多变和边缘模糊等挑战,而且在分割性能上具有更为出色的表现,为医学图像处理领域的进一步研究和临床应用提供了有力支持。
