Automatic Respiratory Phase Identification and Mismatch Correction for Free-breathing PET/CT

Objectives: In normal free-breathing PET/CT imaging protocol, a single CT image of the patient is obtained before PET scan and used for attenuation correction of PET data. A CT scanner with long axial field of view and fast rotation speed is able to capture the patient body at one respiratory phase even with free breathing during acquisition. However, since breath-holding is infeasible during PET scan, the single CT image is inadequate for attenuation correction of the whole PET data. The aim of this study is to design an automatic workflow to correct mismatch of PET and CT data obtained in free-breathing scans of different tracers. Methods: First, the PET data was divided into four equal-count respiratory frames based on a previously developed data-driven respiratory gating method. The attenuation map from CT scan was modified by filling lung region with attenuation coefficient of soft tissue before use in attenuation correction of the gated PET data to reduce activity-attenuation mismatch. Second, the respiratory phase of the original CT image was identified automatically. Two regions of interests (ROIs), each containing one of the two lungs and organs below the lung, were segmented from CT image. The mutual information (MI) inside the ROIs of each gated PET reconstruction and the CT image was measured for all gates. Since the left ROI with lung-stomach boundary and the right ROI with lung-liver boundary may have different contrasts for different patients and with different tracer types, MI measurements of two ROIs at all gates were ranked separately. The ROI with largest minimum-maximum difference was used for phase determination, and the gate with maximum MI in this ROI was the respiratory phase of CT image. The automatically determined phase was validated with visual inspection of PET and CT images. Third, gated PET reconstructions generated with the modified attenuation map were used for motion vector fields (MVF) estimation. MVFs from every other gate to the reference gate were estimated using a B-spline based multi-resolution image registration algorithm. The cost function of the algorithm consists of summed square error between two images and smoothness constraint of the MVF. Fourth, the CT image was transformed to every other frame using estimated MVFs to obtain attenuation maps for all respiratory frames. Finally, image reconstruction of gated PET data was repeated using phase-matched gated PET-CT data pairs. Results: This method was applied to