Research On Intelligent Analysis And Processing Technology Of Human Ear CT Images Based On Deep Learning

The intelligent processing and analysis of medical images aims to use automatic technical means to automatically analyze and process medical images,provide reference and help for clinical medical staff,reduce the problems of misdiagnosis caused by doctors' lack of experience and heavy work,and improve the efficiency of doctors.In recent years,with the rapid development of artificial intelligence technology represented by deep learning,the recognition speed and accuracy of medical images have been greatly improved.Deep learning technology has been initially applied in the automatic analysis of medical images such as skin,brain,chest and abdomen organs.The intelligent processing and analysis of medical images has become one of the most concerned and fastest-growing directions in the field of the Wise Information Technology of 120(WIT120).Ear is an important part of the head and neck of the human body.It plays an important role in the sense of hearing and assisting the body to maintain balance.Ear health is an important factor affecting people's quality of life.Imaging is an important means of modern ear disease diagnosis and treatment technology.Compared with other medical images,the ear structure is complex and precise,the imaging area is extremely small,the individual differences are large,and the sample data is lacking,which brings great challenges to the intelligent processing and analysis of the ear.This paper is aimed at the actual clinical needs of the automated processing and analysis of computed tomography(CT)images of the ear from the Department of Radiology of Beijing Friendship Hospital affiliated to Capital Medical University.With deep learning as the main technical means,the research on the intelligent processing and analysis technology of human ear based on CT image is carried out.The research content of this paper mainly includes the following three parts:(1)A basic framework for developing a normal human ear atlas based on CT volume images is proposed.In clinical diagnosis,treatment and research,establishing an atlas is one of the most important steps in quantitative analysis of medical images.However,for extremely small and precise human ear anatomical structures,the atlas based on CT volume data is less developed currently.Firstly,normalized multi-planar reconstruction of the collected temporal bone CT images is performed.Then voxel-based annotation of 10 key ear structures is carried out at the axial,coronal and sagittal views,including the malleus,incus,stapes,cochlea,vestibule,superior semicircular canal,lateral semicircular canal,posterior semicircular canal,jugular foramen,and internal acoustic meatus.Finally,the volumetric and spatial data of 10 key ear structures are analyzed by descriptive statistics,and a 3D human ear atlas from 64 normal cases of temporal bone CT images is established.The 3D normal human ear atlas created in this paper will provide support and reference for the clinical research of otology imaging,and provide a wealth of anatomical prior knowledge and guidance for the intelligent processing and analysis of the ear CT images.(2)A vestibule segmentation network based on multi-dimensional feature fusion is designed and realized.In view of the extremely small resolution of the vestibule,the multi-scale and spatial position changes of different slices,a method of segmentation of vestibule structure is proposed.And three feature fusion strategies are proposed,which are convolutional feature fusion for different receptive fields,feature channel fusion based on attention mechanism and feature fusion of the decoder-decoder.Based on the three feature fusion strategies,the segmentation accuracy of the vestibule can achieve an average IOU of 88.71%,which is much higher than the comparative state of the art methods.(3)A vestibular deformity recognition method based on transfer learning is proposed.By comparing the impact of two transfer learning strategies on recognition performance,the prior knowledge of vestibular segmentation network is successfully transferred to the vestibular deformity recognition network.The transfer learning strategies include: the first is the transfer learning strategy based on the pre-training model.The trained model is used to initialize the parameters of the vestibular deformity recognition network.The second is the transfer learning strategy based on the segmentation features.The effects of different segmentation feature migration on the recognition performance are tested.The experimental results show that the network recognition performance using the pre-training model fine-tuning is higher than the network recognition performance without fine-tuning.The strategy based on the segmentation feature migration has better recognition performance than the transfer strategy of the pre-training model.