最小描述长度优化下的医学图像统计形状建模

Statistical shape modeling based on minimum description length optimization in medical images

统计形状模型(SSM)是有效的图像处理与分析方法。为了建立模型,需要从形状样本集中提取出具有对应关系的轮廓采样点集合,这是决定模型性能的关键。传统的手动标定这些点集来确保对应关系枯燥耗时且带有主观性,更难以向高维拓展。对形状建立逐层的多尺度参数表示,基于最小描述长度(MDL),在粗尺度上建立反映点对应程度的目标函数并最小化,提出首先确保粗尺度上具有最优意义的点对应,同时在精尺度上使用最便捷的弧长参数函数来确定特征点,完成感兴趣目标的快速统计形状建模,进而统计分析以验证模型性能,为后续图像分割或定量分析打下基础。实验对肌肉骨骼核磁共振成像(MRI)中椎骨、椎间盘以及半月板等具有临床意义的结构建立了统计形状模型,验证了本文方法与手动取点相比具有客观可重复性且更加简洁,与单一尺度下的MDL方法相比时间效率更高。基于此模型的图像分割与基于手动建模的分割相比,误差相当或有所降低。

Statistical shape model （SSM） is an efficient method of image processing and analysis. One key factor in building models is to obtain correspondent landmarks among the whole shape dataset. Traditional manual land-marking is temporally expensive, subjective, boring and prohibitively extensive in dimension. In this paper, a multi-scale parameterization on shapes allows a minimum description length （MDL） based optimization on landmark correspondence in a coarse scale and a most convenient arc parameterization based landmark correspondence in a fine scale. This achieves a fast and accurate SSM building, which is the foundation on following image segmentation and quantitative analysis. In experiments, SSMs are built with vertebral body, intervertebral disc and meniscus shapes extracted from various MRIs respectively. It is testified that the models built with the proposed scheme is not only more repeatable and concise than model baseds on manual landmarking, but also more temporally efficient than model purely based on optimization. The segmentation errors from the proposed method are comparable with or better than those from the manual modeling based segmentation.