高级检索

结合稠密特征映射的CT图像肿瘤分割模型

CT Image Tumor Segmentation Model Combined with Dense Feature Mapping

  • 摘要: 与常规分割对象不同,医学图像中肿瘤组织像素占比小且解剖结构相近于人体狭小组织,不同肿瘤之间差异度不明显,这导致常规分割方法对肿瘤的分割效果低于期望值.因此,为了增强肿瘤特征传递的有效性,提出一种结合高低维稠密特征映射的肿瘤分割模型.首先,模型采用特征3维映射技术改进网络参数,将CT图像聚合成3维序列结构进行硬阈值3维变换,从而建立特征连接并减少不可逆初始特征丢失现象.然后,构建融合特征映射的稠密卷积网络,使用SELU代替ReLU激活函数,激活网络并提升网络优化度,引入负数部分参数避免“死特征”出现,并在每个稠密块后增加一层最大池化层抽象图像特征,减少时间、空间资本消耗.最后,采用特征复现方法进行特征重建,融合通道特征、空间特征提升特征表达能力.实验采用山东省千佛山医院提供的CT图像数据集,在TensorFlow环境下将模型与U-Net等分割模型进行对比,并对模型进行了消融实验.实验结果表明,该模型有效地提升了肿瘤分割的准确度,与已有经典模型相比,在均像素精度、均交并比等性能指标上均取得了更好的效果.

     

    Abstract: Different from the conventional segmentation object,the proportion of tumor tissues in medical images is small and the anatomical structure is similar to that of human narrow tissues.Furthermore,the un-clear difference between tumors leads to the poor segmentation performance compared with the expected value.Therefore,in order to enhance the effectiveness of tumor feature transfer,a tumor segmentation model combined with high/low dimensional dense feature mapping is proposed.Firstly,the feature three-dimen-sional(3D)mapping is used to optimize the network parameters,the CT images are aggregated into a 3D sequence structure for hard threshold 3D transformation,for establishing the feature connections and reduc-ing irreversible initial feature loss.Then,a dense convolutional network is built and ReLU is replaced by SELU activation function,in order to activate the network and introduce negative parameters to avoid“dead features”.In addition,a layer of maximum pooling layer is also added after each dense block to reduce the capital consumption of time and space.Finally,via fusing the channel and spatial features,the feature emersion method is adopted for feature reconstruction,which improves the ability of feature expression obvi-ously.The classic segmentation methods such as U-Net are compared under the TensorFlow environment,and ablation experiments are performed.Experiments on CT image dataset provided by Shandong Provincal Qianfoshan Hospital show that proposed model effectively improves the accuracy of tumor segmentation.Compared with the existing classical models,proposed model achieves better performance in terms of aver-age pixel accuracy and average intersection ratio.

     

/

返回文章
返回