A harmonization strategy for deep progressive reconstruction of PET/CT

Introduction: Deep Learning (DL) techniques exhibit immense potential in PET imaging applications. Substantial efforts have been dedicated to developing DL-based PET reconstruction algorithms. Studies have demonstrated the capacity of DL-based PET reconstruction to enhance image contrast, decrease image noise, thereby improving lesion detection, reducing radiation dose, and expediting acquisition time. However, DL reconstruction presents a challenge in the harmonization of the standardized uptake value (SUV), which may be susceptible to variations in training datasets. It is imperative to investigate how discrepancies between the structure of NEMA image quality phantom and human anatomy may impact DL reconstruction algorithms. Thus, our study aims to examine the influence of DL-based PET reconstruction on harmonization strategies, utilizing both NEMA image quality phantom and patients' data.

Methods: A NEMA image quality phantom was meticulously prepared in accordance with EARL specifications. Seven patients with FDG avid lesions underwent PET/CT examinations were retrospectively enrolled. All imaging procedures were conducted using a United Imaging uMI 780 PET/CT scanner reconstructed with ordered subset expectation maximization (OSEM) using 2 iterations and 20 subsets and deep progressive reconstruction (DPR). Initially, a 3D Gaussian filter with various full-width half-maximization (FWHM) settings ranging from 1 to 5 mm was applied to the OSEM-reconstructed phantom images. The recovery contrast of the maximal activity (RCmax) within the spheres was quantified using a in-house MATLAB-based software. FWHM configurations that aligned RCmax with EARL 2 specifications were identified and chosen for subsequent applications to patient images. In the patient images, FDG avid lesions were delineated using MOOSE. Maximal standardized uptake values (SUVmax) of the lesions were measured in EARL-compliant OSEM-reconstructed images.Subsequently, the lesion segmentation masks were transferred to DPR reconstructions with varying Gaussian FWHM settings. The SUVmax of DPR images was measured and compared.

Results: A 3 mm FWHM setting was identified as meeting EARL compliance for the OSEM reconstruction. Regarding DPR reconstruction, the 3 mm FWHM setting ensured RCmax consistency with EARL-compliant OSEM reconstruction for spheres larger than 17 mm in diameter. However, spheres smaller than 17 mm demonstrated more consistent RCmax between DPR and EARL-compliant OSEM reconstructions when using a 4 mm FWHM setting.In the patient study, a total of 118 FDG avid lesions ranging in volume from 235 to 161154 mm3 were analyzed. The SUVmax measured in DPR reconstructions employing 3 mm and 4 mm FWHM Gaussian filters exhibited the closest agreement with EARL-compliant OSEM reconstruction. The ratio of SUVmax between DPR and OSEM reconstructions was calculated and depicted in Fig 1b. The mean and 95% confidence interval (95% CI) of the SUVmax ratio were 0.98 [0.85, 1.11] for DPR with a 4 mm FWHM post-filtering, displaying a closer alignment to a ratio of 1 and a narrower 95% CI range compared to DPR with 3 mm FWHM Gaussian filtering (1.18 [0.88, 1.48]). Additionally, similar to the phantom results, there was a trend of an increasing SUVmax ratio with smaller lesion volumes, albeit with a less pronounced non-linearity than observed in the phantom study.

Conclusions: Our findings suggest that harmonizing DL-based PET reconstruction is indeed achievable. According to our data, employing a 4 mm FWHM Gaussian post-filter enables DPR-reconstructed images to align with EARL 2 specifications. Furthermore, our study offers a valuable strategy utilizing both phantom and patient data, providing a robust protocol for harmonization. This approach effectively mitigates the impact of non-linear DL-based reconstruction on the phantom and patient data, ensuring more reliable and consistent results across imaging scenarios.

• Download figure
• Open in new tab
• Download powerpoint