融合候选区域提取与SSAE深度特征学习的心脏MR图像左心室检测

Combining Region Proposals and Deep SSAE Learnt Features for Detecting Left Ventricle in Cardiac MR Images

左心室检测在计算机辅助心脏MR图像诊断方面具有重要价值,针对由于成像质量、部分容积效应、目标复杂多变等因素影响,导致左心室自动检测准确度较低的问题,提出一种融合候选区域提取与栈式稀疏自编码器(SSAE)深度特征学习的心脏MR图像左心室检测方法.在候选区域提取阶段,先用超像素算法产生初始区域,然后对SSAE学习到的深度特征采用层次聚类算法生成候选区域;在检测阶段,先使用SSAE提取候选区域的深度特征,然后训练SVM分类器对候选区域进行分类,并使用难分负样本挖掘算法对模型进行调节.对心脏图谱数据集左心室目标检测的实验结果表明,相对于手工特征及基于候选区域等方法,该方法取得了有竞争力的检测精度.

Automatic detection of left ventricle(LV)is an important step for further analyzing cardiac MR images.However,due to the image acquisition,partial volume effect,low resolution and high similarity to the surroundings,it is a challenging task for improving LV detection accuracy.In this paper an automatic detection method is proposed by combining region proposals and deep Stacked Sparse Auto-encoder(SSAE)learnt features.It consists of two components:1)At the stage of proposing candidate regions,a superpixel algorithm is firstly adopted to generate initial regions,then a hierarchical clustering algorithm using deep SSAE learnt feature is employed to make the final candidates;2)At the stage of detection,a SSAE network is used to extract deep feature of the resulting candidates,and the learnt feature is used to train a linear C-SVM classifier.Furthermore,a hard negative mining strategy is added for tuning the model adaptive to the sample imbalance problem.Experimental results of left ventricle detection on the Cardiac Atlas Project(CAP)data set show that,compared to the representative hand-crafted or region proposal based methods,the proposed method achieves competitive results.