基于3D U-Net 实现人体耳软骨MRI图像的解剖结构分割

自体肋软骨雕刻法是目前治疗先天性小儿畸形的临床标准疗法,而耳软骨组织工程和3D生物打印是有前景的治疗方案。可是,这些治疗方案的核心—(复合物)支架构造缺乏基于医学图像的耳软骨自动分割方法。基于3D U-Net提出改进的网络模型,能够实现MRI图像的人体耳软骨解剖结构的自动分割。该网络模型结合残差结构和多尺度融合等设计,在减少网络参数量的同时实现12个耳软骨解剖结构的精确分割。首先,使用超短回波时间(UTE)序列采集40名志愿者单侧外耳的MRI图像;然后,对所采集的图像进行预处理、耳软骨和多解剖结构手动标注;接下来,划分数据集训练改进的3D U-Net模型,其中32例数据作为训练集、4例为验证集、4例为测试集;最后,使用三维全连接条件随机场对网络输出结果进行后处理。模型经过10折交叉验证后,耳软骨12个解剖结构的自动分割结果的平均Dice相似度系数(DSC)和平均95%豪斯多夫距离(HD95)分别为0.818和1.917,相比于使用基础的3D U-Net模型,DSC指标分别提高6.0%,HD95指标降低了3.186,其中耳软骨关键结构耳轮和对耳轮的DSC指标达到了0.907和0.901。实验结果表明,所提出的深度学习方法与专家手动标注两者之间的结果非常接近。在临床应用中,根据患者健侧UTE核磁图像,本研究提出的方法既可以为现有自体肋软骨雕刻法快速、自动生成三维个性化雕刻模板,也可以为组织工程或者3D生物打印技术构建耳软骨复合物支架提供高质量的可打印模型。

Anatomical Structure Segmentation of Human Auricular Cartilage MRI Images Based on 3D U-Net

The method of costal cartilage carving is currently the clinical standard treatment of microtia, and auricular cartilage tissue engineering and 3D bioprinting are promising approaches. However, there has been a lack of automatic auricular cartilage segmentation based on medical images that is the crucial and fundamental issue of the treatments. In this study, an improved network based on 3D U-Net was proposed to automatically segment the anatomical structures of human auricular cartilage on MRI images. The proposed network combined the residual structure and multi-scale fusion design to reduce the number of network parameters and achieved an accurate segmentation of 12 auricular cartilage anatomical structures. Firstly, the Ultra-short Echo Time (UTE) sequence was applied to collect MRI images of the unilateral external auricular of 40 volunteers; secondly, manual segmentation of both auricular cartilage and multiple anatomical structures were performed on the preprocessed images of each volunteer; next, the images were divided into the training dataset of 32 images, the validation dataset of 4 images, and the testing dataset of 4 images to train the improved 3D U-Net model; finally, the 3D fully connected conditional random field was used to post-process the output of the proposed network. Ten-fold cross-validation was performed on the model, and the averaged Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95) of the automatic segmentation results of the 12 structures were 0.818 and 1.917, respectively. Compared with the basic 3D U-Net model, DSC was increased by 6.0% and HD95 was decreased by 3.186. Especially, the DSC of the key structure, helix and the antihelix, were 0.907 and 0.901, respectively. The experimental results showed that the segmentation results of the proposed method were very close to the manual annotations by experts. 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.