Research On CT Image Noise Identification And Application Based On Dense Net

It is inevitably for CT images to be affected by noise in the process of imaging,transmission,storage and use,and noise may cause doctors to misjudge CT images.Correctly identifying noise types and noise intensity in CT images can not only find out the source of noise but also reduce the blindness and complexity of image denoising.However,most of the existing noise identification algorithms are based on the assumptions that there is only one single type of high-intensity noise in the image,and few researchers study the common low-intensity mixed noises.As for the low-intensity mixed noises,the main work of this paper is as follows:(1)At present,there is no data set for noise images.Based on the data set of CT images of TCGA-LUAD lung cancer,this paper constructed a small data set containing a total of 1200 low-intensity mixed noise images,and the data set includes 8 categories,such as Gaussian noise,Slat & Pepper noise and mixed noise.(2)The paper proposed an image noise identification algorithm based on DenseNet.The algorithm chooses DenseNet which has been pre-trained on ImageNet.The pre-trained DenseNet can efficiently recognize the features of images,such as contour edges,and has a lot of advantages,such as high anti-overfitting,so it is suitable for training small data sets.The algorithm firstly carries out a series of preprocessing on the data set of noise images and extracts the visual statistical feature map of the noise images.Then the fine-tunes are conducted to the network structure of DenseNet.Finally,the fine-tuned DenseNet is used to train the visual statistical features.After training,the network can accurately identify the low-intensity Salt & Pepper noises,Gaussian noises as well as mixed noised in the images.(3)When the image noises are identified,an appropriate denoising algorithm can be selected for denoising.Taking the denoising convolutional neural network DnCNN as an example,this paper constructs 8 types of corresponding denoising convolutional neural networks for 8 kinds of different noises,and compares it with other denoising algorithms,so as to illustrate the application and advantages of image noise recognition in the subsequent denoising process.