Validation of the lesion quantification of a learning-based PET image filter

Introduction: To achieve a favorable visual effect and attain a balance between high-frequency details and the noise, post-reconstruction filtering has been widely adopted in the processing of PET images. Recently, more sophiscated image processing algorithms have been proposed to enhance the image quality while preserving the details of diagnostic significance. Working on the image domain, these algorithms function like post-reconstruction filtering. However, in contrast to traditional Gaussian fitlering, the impact of these algorithms on the quantification metric SUV is not linear, posing a threat to the reliability in clinical scenarios. Such a deep-learning based post-reconstruction filtering method, HYPER DLR (abbreviated as DLR), has obtrained the FDA clearance and the image denoising ability has been validated. DLR takes the input from a reconstructed PET image, and through neural network inference, generates a low-noise, high contrast image. Trained with actual clinical data to capture the morphorlogy-dependent noise property, the DLR algorithm may not be suitable to be evaluated using the traditional methods involving a geometric phantom, which is an outlier from the distribution of the original triaining set. In this research, we compared the quantitative performance of the DLR algorithm with the harmonized OSEM images using SUV_max as a reference, and validated the quantitative accuracy of this algorithm on a small cohort.

Methods: The evaluation of the quantitative performance was first conducted according to the standard EARL procedure, where a NEMA IQ phantom filled with 10:1 contrast ratio hot spheres was scanned and reconstructed. Gaussian filters with different FWHM were applied, and the parameter that allows all recovery coefficients corresponding to the maximum value (RC_max) to fall into the range suggested by EARL2 was adopted. Patients whose [¹⁸F]FDG PET scan showed more than five metabolic-avid lesions were included based on a retrospective analysis in one center over a period of one day. The images were recontructed with the standard OSEM algorithm (3 iterations, 20 subsets, with TOF, and PSF) without applying the post-reconstruction Gaussian filtering. A set of filtered images were reconstructed adopting the same paramters, and incorporating the DLR filtering. Gaussian filters were then applied to both sets of images, obatining the 'OSEM' and 'DLR' images. The images were segmented using an automatic pipeline (the 'clin_pt_fdg_tomour' model in MOOSE), and the masks for individual lesions were extracted and manually refined. The SUV_max and the volume corresponding to each lesion mask were calculated.

Results: According to the phantom, the OSEM algorithm requires a Gaussian filter of 4.5 mm FWHM to conform to the EARL standard, while DLR requires a FWHM of 4.0 mm. Such filters were applied to the corresponding images. The SUV_max of the DLR group is slightly lower than the SUV_max group, and the trend is more significant in smaller lesions. Across different SUV range, the relationship is linear, but with an offset. This is also verified with a Bland-Altman plot.

Conclusions: The quantitative accuracy of the DLR algorithm has been evaluated, and the SUVmax of the DLR algorithm, albeit applying the harmonization procedure, is still slightly lower than the validated OSEM algorithm. In the future, the Gaussian filter should be systematically adjusted to allow the actual convergence to the actual range of the reference method.

• Download figure
• Open in new tab
• Download powerpoint