Research And Application Of Deep Convolutional Neural Network In Segmentation Of Dermoscopic Images And Whole Breast Ultrasound Images

Accurate analysis of medical images is crucial for clinical diagnosis.However,the results of analysis will be slightly different due to the subjective judgment of different doctors.At the same time,the lack of medical resources in China will increase the burden on doctors,and the diagnosis process will be time-consuming and labor-intensive.Therefore,it is urgent to propose computer aided diagnosis methods to alleviate many limitations of manual analysis of medical images.In recent years,with the development of hardware devices and neural network theory,deep learning has achieved amazing performance in many fields including medical image analysis.Compared with traditional computer aided diagnosis methods,deep learning has many advantages,such as: no need for manual interaction,no feature selection;can directly input images of any size to complete end-to-end network training;no complicated post-processing steps are required.However,deep learning applications are still full of challenges in the field of medical image segmentation.For example,when extracting features,it is impossible to distinguish between useful features and redundant features.When the image features are extracted from the network and then restored to the original image,the spatial information of the image is lost.Labeling the data makes the network prone to over-fitting and the like.Aiming at the many challenges in the field of medical image segmentation,this paper takes the dermoscopic image segmentation and two dimensional whole breast ultrasound image as the object,and separately divides the skin lesions automatically and divides the two dimensional whole breast ultrasound image anatomical layer.End-to-end network training is achieved based on the network model of the codec framework.The research results of this paper mainly include the following two aspects:On one hand,automated segmentation of melanoma based on the ISBI Skin Lesion Challenge Dataset of 2016 and 2017.In this paper,a skin segmentation network based on codec framework is designed.By improving the upsampling operation,the method of how the decoding module restores the high resolution of the original image is explored.The decoding module mainly includes three parts: a residual convolution unit,a chain residual pooling layer,and a dense deconvolutional layer.The same resolution of the codec module uses the jump connection to alleviate the problem of gradient disappearance,while using depth supervision to accelerate network convergence.Compared with other existing segmentation methods,the proposed method can enhance the ability of the network to restore the original image resolution and solve the problem of blurred edge of the lesion.In this paper,extensive experiments have been carried out on the skin lesion dataset,and the experimental results demonstrate the effectiveness of the proposed segmentation method.On the other hand,for the anatomical layer segmentation of whole breast ultrasound images,this paper proposes a coding and decoding segmentation network based on attention mechanism,which divides the whole breast ultrasound image into four different tissues from top to bottom: subcutaneous fat layer,breast parenchyma layer,pectoral muscle layer,and the chest wall layer.Specifically,based on the ResNeXt network,the attention module of the spatial dimension and the attention module of the channel dimension are embedded at the same time to enhance the useful features and suppress the useless features.Utilize non-local context modules to increase the range of receptive fields and improve the ability of global context information.For the upsampling operation,this paper proposes a weighted upsampling block to solve the problem of low-level features lacking high-level semantic information.Finally,a large number of experiments were carried out on the whole milk ultrasound image.The experiment proved that the proposed segmentation network effectively solved the problem of anatomical layer segmentation.In summary,based on the dermoscopic image and the whole breast ultrasound image,based on the coding and decoding framework commonly used in image segmentation by deep convolutional neural networks,this paper applies the deep learning method to the problems and challenges in the field of medical image segmentation.A solution to avoid redundant features,extract useful features,and restore high resolution of the original image in upsampling operations.In this paper,a lot of comparative experiments are done on the dataset,which proves the effectiveness of the proposed method.The method proposed in this paper is not limited to dermoscopic images and ultrasound images,but can also be applied to other medical images or natural images.