Evaluation of Segmentation Based Attenuation Correction Methods for PET/MR in the Thorax
H Ai1,2*, T Pan1,2, (1) The University of Texas MD Anderson Cancer Center, Houston, TX, (2) The University of Texas Graduate School of Biomedical Science, Houston, TXSU-E-I-86 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
To compare and estimate the accuracy of different segmentation-based approaches for attenuation correction of the thoracic PET data in PET/MR.
PET/CT data were retrospectively collected from 26 patients who underwent whole-body PET/CT scan. CT images of each patient were segmented into five categories, including air, lung, fat, soft tissue and bone, with a fast 3D fuzzy c-means clustering algorithm. The mean CT numbers of different segments for each patient were measured. For the 20 patients who have legions found in the thorax, 4 sets of CTAC images were created according to the segmentation results, simulating the attenuation maps with four different approaches: 5-class segmentation assigned with individually measured CT number (SFLAB) and with mean CT number of the 26 patients (SFLAB mean), 4-class with mean CT number (SFLA mean) and 3-class with mean CT number (SLA mean). Attenuation corrected PET images were generated with these CTAC images and compared to the original attenuation corrected PET images using the maximum SUV of 35 legions identified in the thorax, both inside and outside of the lung.
The mean/standard deviation of the measured CT number in the 26 patients are -703/58 (lung), -105/6 (fat), 37/8 (soft tissue) and 483/63 (bone), with air CT number = -1000. The maximum SUV differences in the 35 thoracic legions are -0.7%±3.2% (SFLAB), -1.0%±5.0% (SFLAB mean), -4.0%±5.7% (SFLA mean) and 5.7%±6.3% (SLA mean), with p-values being 0.2069, 0.2303, 1.752e10-4 and 5.737e10-6, respectively.
5-class segmentation methods achieved the highest quantification accuracy of PET activity in the thorax. 4-class method tends to under-estimate while 3-class method tends to over-estimate the maximum SUV. Despite the simplicity of the approaches, segmentation-based methods can achieve reasonably high accuracy in the quantification of thoracic PET data.