Application Of Deep Learning In Temporal Bone CT

Objective The temporal bone is one of the most complex bony structures because of the intricate internal structure.The facial nerve,ossicle chain and cochlear are closely related.The temporal bone CT examination is one of the most important examinations in the diagnosis and treatment of ear diseases,automatic segmentation and diagnosis can help us to make a diagnosis and achieve precise treatment.Our aim of this study was to assess the utility of deep learning analysis in diagnosing cochlear malformation and automatic segmentation of cochlear,facial nerve,membranous labyrinth,internal auditory canal and ossicular chain on temporal bone CT images.Methods In this study,temporal bone CT images were taken from 1800 patients undergoing cochlear implantation in the First Affiliated Hospital of Guangxi Medical University from 2003 to 2021.Preoperative imaging diagnosis was completed by 3 associate chief physician from otorhinolaryngology head and neck surgery,and manual segmentation was completed by an associate chief otologist with rich experience in temporal bone CT imaging reading.A chief doctor to review and confirm.1.Using 3D U-Net and UNETR models,90 normal temporal bone CTs were trained and 7 normal temporal bone CTs were tested under different hyperparameters to achieve automatic segmentation of facial nerve,membrane labyrinth,internal auditory canal and ossicular chain.2.97 cases of normal cochlear were trained by 3D U-NET and UNETR models,and 10 cases of normal cochlear and 5 cases of cochlear malformation were used as test sets for automatic segmentation.3.The temporal bone CT images of 373 malformed cochlea and 390 normal cochlear were preprocessed with three different kinds of images in four deep learning models of Res Net10,Res Net50,SE-Res Net 50 and Dense Net121.The test data set was then diagnosed by physicians of different seniority and compared with Res Net10,Res Net50,SE-Res Net 50 and Dense Net121 network models for automatic diagnosis of cochlear malformation.Results 1.UNETR model can achieve better segmentation performance than3 D U-Net.Epoch 200 and batch size 2 can achieve the best performance.The DSC of facial nerve,membranous labyrinth,inner auditory canal and ossicular chain are 0.65,0.91,0.86 and 0.78,respectively.2.Using UNETR model(batch size=1),the average DSC of normal cochlea automatic segmentation testing set was 0.92 and cochlear malformation testing set was 0.93,which was superior to 3D U-Net model.3.In the automatic diagnosis of cochlear malformation,the training and modeling method of 2 2D axial CT images in Res Net10 obtained the best diagnostic performance,with an AUC of 0.96.There was less human intervention in 3D image preprocessing,and the diagnosis time was the shortest.The diagnostic performance of deep learning is comparable to that of senior otologists.Conclusions The establishment of automatic diagnosis and segmentation model through deep learning may become a routine means to assist in diagnosis of cochlear malformation,which can reduce the missed diagnosis rate of cochlear malformation and improve the clinical diagnostic efficiency.The automatic segmentation of temporal bone anatomy is helpful to assist the otologist to make a reasonable planning of surgery and select the appropriate surgical method.