Encrypted login | home

Program Information

Multi-Atlas and Learning Based Segmentation of Head and Neck Normal Structures From Multi-Parametric MRI

no image available
H Veeraraghavan

H Veeraraghavan*, N Tyagi , M Hunt , N Lee , J Deasy , Memorial Sloan Kettering Cancer Center, New York, NY


SU-F-303-16 (Sunday, July 12, 2015) 4:00 PM - 6:00 PM Room: 303

Purpose: To generate automatic segmentation of head and neck normal structures from multi-parametric MR Dixon images.

Materials and Method: We present a multi-atlas based registration combined with machine learning-based segmentation of head and neck structures from MR T1 FFE based mdixon images. All the images were acquired using a 3T Phillips MR scanner. 6 patients; 3 in atlas and 3 in testing were used. The individual mdixon images were registered to corresponding images from the multi-atlas using affine and B-spline deformable registration using the open-source software Plastimatch. Second, the best-aligned image pairs were automatically extracted through automatic landmark generation and matching. Scale Invariant Feature Transform (SIFT) features were used to generate the landmarks. The segmentation labels were propagated from the best matching atlas. A random forest (RF) classifier trained with K=10 fold cross validation using the MR mdixon and Haralick textures computed on the same images from the multi-atlas refined the segmentation. Finally, the generated segmentations were smoothed using Markov Random Field and morphological post-processing.

Results: The patients used in the analysis displayed anatomical variations owing to dental implants and disease. We evaluated the segmentations generated by our method by computing dice overlap scores with manually generated segmentations. Our method resulted in an accuracy ranging between 0.5 to 0.73 for the various structures, namely, bone, right and left parotid, right and left submandibular glands. The algorithm selected registrations closely agreed with the visual comparison of the image registrations.

Conclusions: We developed a fully automatic method for normal structures segmentation that combines multi-atlas based registration with machine learning from multi parametric MRI. Our method quantifies the accuracy of registrations by using automatic landmark extraction. Accurate, automatic volumetric segmentation of normal structures is essential for MR-based treatment planning.

Contact Email: