基于神经网络和体素模板的骨骼受损类型判别

Discrimination of Fracture Types Based on Neural Network and Voxelized Bone Template

针对骨骼受损类型复杂多样、难以自动判别的问题,提出一种基于神经网络和体素模板的骨骼受损类型自动判别方法.首先构建一种区域分割且规则化的体素模板,以有效地表征形态结构不规则的骨骼受损区域;然后建立一种受损骨骼与体素模板之间的同构映射,用于提取受损区域的体素信息,并依此生成受损类型体素样本库;再结合医学先验知识定义一种受损区域体素间的约束关系,将连续受损区域作为单元,对同类型样本进行组合以扩充样本库;最后设计和训练神经网络模型对骨骼的受损类型进行自动判别.实验中采集352份股骨受损样本,其预测结果与骨科医师的临床诊断结论相比,准确率达97%,且分类准确率、时间性能和所识别的受损类型数目优于现有文献方法,结果表明,该方法能够辅助医生快速、有效地判断患者骨骼的受损类型,为骨折手术中内固定植入物的选取提供理论基础.

In view of the complex and diverse fracture types which are difficult to identify automatically,a novel method is put forward based on a neural network and a voxelized bone template.Firstly,a re-gion-divided and regular voxel template is constructed to effectively indicate the damaged region with ir-regular morphological structure.Secondly,an isomorphic mapping from a damaged bone to the template is established to extract the voxel information of its fracture region,and a database of fracture types is created accordingly.Next,a kind of constraint relationship between voxels in the damaged area is defined based on the medical prior knowledge,and the continuous damaged areas are considered as units to combine samplesof the same type for augmenting the samples.Finally,a neural network is designed and trained on the aug-mented samples to identify fracture types.The 352 samples of femur fracture are collected in the experiment,and the coincident rate between the predicted result and the clinical diagnosis of orthopedists is 97%.The classification accuracy,time performance and the number of damage types identified are better than the ex-isting methods.Experimental results show that the proposed method can assist doctors to quickly and effec-tively determine the fracture types in patients,and provides a theoretical basis for the selection of internal fixation plate in fracture surgery.