Cluster-based segmentation of dual-echo ultra-short echo time images for PET/MR bone localization
Gaspar Delso1, Konstantinos Zeimpekis2, Michael Carl3
1GE Healthcare, Waukesha, WI, 53186, USA. gaspar.delso@usz.ch.
Background:
Magnetic resonance (MR)-based attenuation correction is a critical component of integrated positron emission tomography (PET)/MR scanners. It is generally achieved by segmenting MR images into tissue classes with known attenuation properties (e.g., bone, fat, soft tissue, lung, air). Ultra-short echo time (UTE) have been proposed in the past to locate bone tissue. In this study, tri-modality computed tomography data was used to develop an improved algorithm for the localization of bone in the head and neck.
Methods:
Twenty patients were scanned using a tri-modality setup. A UTE acquisition with 22-cm transaxial and 24-cm axial field of view was acquired, with a resolution of 1.5 × 1.5 × 2.0 mm(3). The sequence consisted of two echoes (30 μs, 1.7 ms) with a flip angle of 10° and 125-kHz bandwidth. The CT images of all patients were classified by thresholding and used to compute maps of the posterior probability of each tissue class, given a pair of UTE echo values. The Jaccard distance was used to compare with CT the bone masks obtained when using this information to segment the UTE datasets.
Results:
The results show the desired bony structures as a cluster pattern in the space of dual-echo measurements. The clusters obtained for the tissue classes are strongly overlapped, indicating that the MR data will not, regardless of the chosen space partition, be able to completely differentiate the bony and soft structures. The classification obtained by maximizing the posterior probability compared well to previously published methods, providing a more intuitive and robust choice of the final classification threshold. The distance between MR- and CT-based bone masks was 59% on average (0% being a perfect match), compared to 76% and 69% for two previously published methods.
Conclusions:
The study of tri-modality datasets shows that improved bone tissue classification can be achieved by estimating maps of the posterior probability of voxels belonging to a particular tissue class, given a measured pair of UTE echoes.
