Research On Segmentation Algorithm Of Human Ear Cartilage MRI Image Based On Deep Learning

Microtia is a common congenital maxillofacial deformity,which is usually characterized by auricle hypoplasia,and auricle reconstruction is a plastic surgery method for the treatment of this disease.The costal cartilage carving method is currently the clinical standard treatment of microtia,and auricular cartilage tissue engineering and 3D bioprinting are promising approaches.At present,the key point of these treatments,(composite)scaffold construction,lacks effective automatic auricular cartilage segmentation methods.To solve the problem of accurate segmentation of auricular cartilage image,this paper proposes three parts of work,including manual anatomical structure segmentation strategy of auricular cartilage based on Ultra-short Echo Time(UTE)MRI image,auricular cartilage and anatomical structure segmentation based on improving 3D U-Net network,auricular cartilage segmentation based on the cascade network of registration and segmentation.In the first part of the work,first of all,we preprocess and manually segment the auricular cartilage from the UTE images of 40 healthy volunteers;Then,according to the anatomical and morphological characteristics,the segmentation strategy of auricular cartilage anatomical structure is proposed and formulated;Finally,according to the established segmentation steps,the ear cartilage was divided into 12 structures,namely,the antihelix,the crura anthelicis,the fossa triangular,the helix,the crus of helix,the cymba conchae,the scapha,the tragus,the antitragus,the intertragic notch,the external auditory canal,the cavum conchae.Statistical analysis results show that the proposed segmentation method can repeatedly segment 12 structures based on UTE images of auricular cartilage.In the second part of the work,firstly,for the UTE image segmentation of auricular cartilage,a variety of improved designs are proposed,including deepening model depth,introducing residual module in the coding layer,convolution substitution pooling and multi-scale fusion;Then,the images are divided into the training dataset of 32 images,the validation dataset of 4 images,and the testing dataset of 4 images;Finally,auricular cartilage segmentation is realized by using the improved network of manual segmentation ear cartilage label supervision training.Cross-validation was performed on the model,and the average Dice similarity coefficient(DSC),95%Hausdorff distance(HD95)and volumetric similarity(VS)of the automatic segmentation results of the auricular cartilage were 0.872,0.854 and 0.983,respectively.Compared with the classical 3D U-Net model,the DSC,HD95 and VS of our proposed model are increased by 2.6%,decreased by 0.178 and increased by 5.9%respectively.In addition,in order to further segment the auricular cartilage anatomical structure,the manual segmentation of auricular cartilage anatomical structure label is used to supervise the training network.The loss function is improved to a weighted dice loss function for auricular cartilage structure,and a three-dimensional fully connected conditional random field is added to post-process the network output.The results showed that the average DSC,HD95 and VS of the automatic segmentation results of the 12 structures were 0.813,2.085 and 0.920,respectively.Compared with the classical 3D U-Net model,the DSC,HD95 and VS of our proposed model are increased by 6.0%,decreased by 1.551 and increased by 8.3% respectively.In the third part,in order to reduce the dependence of network training on manual labels,a three-level cascade network is designed to segment auricular cartilage.Firstly,the images are divided into the training dataset of 36 images and the testing dataset of 4 images;Then,in order to deform the referenced labels,the proposed model performs the non-rigid registration using two cascaded and unsupervised networks of deformable registration to estimate global and local deformation vector fields(DVFs),respectively;Then,by warping these successively concatenated DVFs to the reference labels,it provides the training dataset with comparably precise cartilage annotations;Finally,these automatically generated labels are used to supervise the training of the segmentation network with the attention gate to further improve the segmentation accuracy.When only one reference annotation was used,the results showed that the DSC,HD95 and VS of the automatic segmentation results were 0.837,1.078 and 0.969,respectively.In this paper,we propose a method for manual anatomical segmentation of auricular cartilage;An improved 3D U-Net is proposed to segment auricular cartilage and its anatomical structure,which makes the segmentation result closer to the manual label;Finally,a cascade framework of registration and segmentation is proposed,which can achieve high-quality segmentation of auricular cartilage with only a few manual labels.In clinical applications,based on the UTE image of the unilateral or parental auricle,the proposed method can quickly and automatically generate a 3D personalized craving template for the scaffold reconstruction with autologous costochondral cartilage and provide high-quality printable model for tissue engineering or 3D bioprinting technology to construct the composite scaffold with detailed auricular cartilage shape.