基于随机森林算法的小鼠micro-CT影像中骨骼关节特征点定位

随着小动物成像技术的发展,技术人员每天需要处理的小动物影像数量急剧增长,这使得自动化的小动物图像分析方法成为迫切的需求。在小鼠图像分析方面,小鼠灵活多变的身体姿态给自动化的图像分析带来困难。基于随机森林算法实现小鼠micro-CT图像中骨骼关节点的自动定位,为解决小鼠影像中身体姿态的自动识别打下基础。该算法主要分3步:先通过分类随机森林算法得到小鼠骨骼关节点的粗定位,再通过回归随机森林算法进一步减小定位误差,最后通过图匹配的方法在备选点中挑选正确位置上的关节点。对49例不同身体姿态的小鼠全身三维micro-CT图像进行测试,全身关节点定位的成功率为98.27%,定位误差的中值为0.68 mm。同时验证联合使用分类与回归随机森林的必要性,并探究训练数据的数量对不同骨关节的识别效果的影响。研究为小鼠micro-CT影像中身体姿态的识别提供一种新方法,为后续的自动化图像配准、图像分割以及自动化图像测量提供重要的定位信息。

Bone Joints Localization in Mouse Micro-CT Images Using Random Forests Algorithm

Along with the rapid development of imaging techniques for small animals, more and more images obtained from small animals need to be analyzed per day, therefore automated image analysis method has become an urgent demand. For mice images, the significant inter-subject posture variations become a major difficulty for automated image analysis. In this paper, an automatic bone joint localization method was developed for mouse micro-CT images, so as to help with posture identification of mouse body. The proposed method was composed of three steps: (1) classification random forests for rough joint localization, (2) aggregating the results of classification through regression forest, and (3) picking up landmarks in the right position by the mapping graph. The method achieved automatic bone joint localization for 49 test images of different body postures. The median localization error of the whole body CT images was 0.68 mm. The success rate of localization was 98.27%. We also demonstrated the necessity of combining classification and regression random forest and discussed the influence on localization with different number of training data. With this new method for mouse micro-CT posture identification was expected to provide helpful information for the subsequent image registration, segmentation and measurements.