Deep-learning-based image warping to reduce spatial misregistration of PET and CT in whole-body scans with 18F-FDG and 64Cu-DOTATATE

Introduction: In PET/CT, patients should optimally have similar positioning of their anatomical organs between the PET and CT portions, yet misregistration (MREG) occurs. We have noted a higher incidence of MREG in ⁶⁴Cu-DOTATATE PET/CT, compared to whole-body ¹⁸F-FDG PET/CT, presumably due to the DOTATATE scans’ longer duration of PET acquisition and higher likelihood of respiratory misregistration and motion. As previously described, a deep learning approach can warp CT to improve the registration (REG). This uses two convolutional neural networks (CNN), trained by humans, to generate displacement vector fields (DVF) that relate PET to CT, and generates a PET image (CNN-PET) with a reduced level of motion-related attenuation artifacts, for example creating accurate standardized uptake values (SUV). However, CNN-PET is not registered with the acquired CT image. This work proposes and evaluates a prototype warping-based registration method (WARP) for registering CNN-PET to CT.

Methods: The WARP method used the DVF to linearly interpolate or warp the CNN-PET image, creating a new image WARP-PET that in principle had accurate SUV and was well registered with CT. To test the method, following IRB approval, we retrospectively defined a study with 20 oncological patients referred for ¹⁸F-FDG PET/CT and 10 neuroendocrine cancer patients referred for ⁶⁴Cu-DOTATATE PET/CT. Scan data were acquired in a Biograph-mCT Flow PET/CT scanner with a fast spiral acquisition for CT and continuous bed motion for PET. ¹⁸F-FDG and ⁶⁴Cu-DOTATATE PET/CT scan durations were approximately 15 and 35 minutes, respectively. The longer time for the DOTA acquisition was used because of the isotope’s low positron yield. For each patient, PET and WARP-PET images were presented to three experienced nuclear medicine physicians, who did not know which image was which. The physicians identified lesions with elevated tracer uptake and in each case characterized the lesion’s alignment with CT as good, acceptable, poor, or cannot tell, and recorded their answers in a spread sheet. The physicians were asked the same question in regard to the alignment of kidneys, liver dome and (in the case of the DOTATATE tracer) spleen. Answers were used to characterize the frequency of MREG in PET and WARP-PET. Finally, maximum SUV of the selected lesions were recorded. Two hypotheses were tested: first, that the incidence of MREG was highest in PET (DOTATATE) not corrected, next highest in PET (FDG) not corrected, and lowest in WARP-PET (FDG or DOTATATE); and, second, that for lesions within 3 cm of the diaphragm, SUV was on average elevated in the case of WARP-PET, compared to PET.

Results: The magnitude of DVF near the diaphragm confirmed the impression of relatively larger magnitude of MREG in DOTATATE studies compared to FDG. More than half of the DOTA studies exhibited a superior-inferior (SI) shift of the liver and spleen by as much as 3 cm. Manifest attenuation artifacts in that territory were reduced by the use of CNN-PET. Smaller left-to-right displacements were noticed in some lesions and in the entire habitus, suggestive of gross motion between the times of the CT and PET scans. The MIP comparison of PET and WARP-PET indicated that the two were in general well aligned, while attenuation artifacts were also greatly reduced in WARP-PET compared to PET. SUV were elevated in WARP-PET images near the diaphragm, compared to PET. In some cases the warping procedure resulted in reduced image sharpness. It was noted that the the MREG reduction was sometimes more pronounced in the liver's superior aspect than in the inferior aspect.

Conclusions: The impression that longer-duration PET scans have a greater incidence of misregistration artifacts was supported by this study. The prototype deep-learning based WARP-PET method reduced these artifacts in whole-body PET/CT, and elevated lesion SUV in the area near the diaphragm.

• Download figure
• Open in new tab
• Download powerpoint