Machine Learning
FIMH '09 Proceedings of the 5th International Conference on Functional Imaging and Modeling of the Heart
ISBI'09 Proceedings of the Sixth IEEE international conference on Symposium on Biomedical Imaging: From Nano to Macro
Spatial decision forests for MS lesion segmentation in multi-channel MR images
MICCAI'10 Proceedings of the 13th international conference on Medical image computing and computer-assisted intervention: Part I
IEEE Transactions on Signal Processing
Hi-index | 0.00 |
This work addresses the challenging problem of segmenting the myocardium in 3D LV echocardiograms by Random Forests (RF). While the RF framework has proven to be a good discriminative classifier for segmentation of 3D echocardiography [1], our hypothesis is that richer features than those traditionally used (Haar etc) need to be employed for accurate segmentation to tackle artifacts in ultrasound images such as missing anatomical boundaries. To address this, we propose two new context rich and shape invariant features, called LoCo and GloCo. The new features impose a local and global constraint on the coupled endocardial and epicardial shape of the left ventricle and use barycentric coordinates to uniquely identify the position of a voxel with respect to a number of landmarks on the epicardial and endocardial border. The landmarks are found using a new measure (COFA) to separate the two boundaries. Experimental results show that the new features provide a smoother segmentation and improve the accuracy compared with a classic RF implementation.