Research On Neck Lymph Node Recognition Algorithm Based On Deep Learning

In medical image recognition,recognition algorithms based on traditional machine learning are usually inseparable from human intervention,and sometimes even require operators to have rich clinical experience and medical prior knowledge.Compared with the traditional machine learning algorithm,the recognition algorithm based on deep learning can automatically extract the high-dimensional semantic features of the image and avoid the performance problems caused by human factors in traditional algorithm.It has become a key research direction in the field of visual computing.Therefore,lesion detection and target recognition in medical images based on deep learning is a cutting-edge research topic.Taking the cervical lymph node as the research object,experiment mainly explores the lymph node recognition algorithm based on deep learning.In view of the imaging characteristics and recognition difficulties of lymph nodes,such as complex pathology,small shape,irregularity and wide distribution,briefly discusses the main work of this paper from the following aspects:(1)In view of the lack of image data and labeling in medical images,based on the limited medical image data to enrich the data set was exploring in this paper.In addition to conventional augmentation,elastic deformation is also one of the most important augmentation strategies.The medical image data of the real environment are simulated by small elastic deformation,so as to allow the neural network to learn the deformation invariance of medical images.A number of experiments and studies have proved that the augmentation strategy is very effective and helps to improve the target recognition index.(2)In order to solve the problem that the current algorithm is easily disturbed by many unrelated organizations,a recognition algorithm based on cascaded full convolution network was proposed,and proposed a strategy of partition and then extraction based on rich medical prior knowledge.Meanwhile,the concept of average pooling of feature blocks was proposed to replace the full connection layer of the traditional three-dimensional classification network,in order to solve the problem that the false positive removal effect is not ideal due to the change of the morphological scale of lymph nodes.It enables the network to provide more accurate discrimination results according to the original spatial information of the samples.(3)In order to solve the current algorithm can't balance the global information to identify the target and avoid the bias to the target position,a semi-random sampling strategy was proposed,which combines the advantages of three-dimensional local location algorithm and three-dimensional global location algorithm,at the same time to achieve effective data enhancement.In addition,an adaptive receptive field recognition network based on attention mechanism was proposed in order to solve the problem of low overall recognition accuracy of conventional network caused by the change of lymph node scale,it dynamically adjusts its receptive field according to different stimuli and fuses multi-scale features to enhance target recognition with obvious scale differences,and achieve better recognition results on the premise of less computational cost.