A Deep Learning model for Partial Volume Correction without Using Prior Knowledge Information

Introduction: The partial volume effect (PVE) is a consequence of the limited spatial resolution of PET scanners. PVE causes the underestimation of intensity values of a particular voxel due to interference of the surrounding background. Partial volume correction (PVC) techniques can be applied to overcome the negative impact of PVE on PET images. A number of popular PVC methods for brain PET, such as Meltzer’s method, Müller-Gärtner (MG), and the geometric transfer matrix methods (GTM), typically require anatomical imaging modality (MRI) as support. This dependence gives rise to a key disadvantage: the need for accurate registration of PET and MR images. This dependency results in inaccurate segmentation, thus resulting in errors in PVC.

Methods: A clinical brain PET dataset consisting of 50 ¹⁸F-FDG, 36 ¹⁸F-Flutemetamol, and 76 ¹⁸F-Fluorodopamine, and their corresponding T1-weighted MR images were enrolled in this study. The Iterative Yang technique was selected as the ground truth PVC technique. A cycle-consistent adversarial network (CycleGAN) was trained to directly map PET images to PVC PET images. Quantitative analysis was performed by calculating structural similarity index metrics (SSIM), root mean squared error (RMSE), and peak signal-to-noise ratio (PSNR). In addition, radiomics analysis was performed by calculating 20 radiomic features within 83 brain regions. Finally, a voxel-wise two-sample t-test comparing the predicted PVC images with original PVC images for each tracer was performed.

Results: The effectiveness of our model for highlighting the anatomical information in predicted PVC is observable. It is worth reminding again that predicted PVC images are synthesized from only PET images in contradiction of reference PVC generated through both MR and PET images. The PSNR varies from the lowest (29.64±1.13 dB) for ¹⁸F-FDG to the highest (36.01±3.26 dB) for ¹⁸F-Flutemetamol. The smallest and largest SSIM was calculated for ¹⁸F-FDG (0.93±0.01) and the¹⁸F-Flutemetamol (0.97±0.01). The relative error average for the kurtosis radiomic feature was 3.32%, 9.39%, and 4.17%, and for the NGLDM_contrast feature relative error was 4.74%, 8.80%, and 7.27%, for ¹⁸F-Flutemetamol, ¹⁸F-Fluorodopamine,and ¹⁸F-FDG respectively. The reference PVC generated from MR + PET images shows MR artifacts/abnormalities propagate into the PVC images, while the deep learning-based PVC images are immune to these artifacts. The results of voxel-wise t-test analysis between predicted and original PVC images show a good correlation between predicted and reference PVC images.

Conclusions: In this work, an end-to-end CycleGAN PVC method was presented and evaluated. Our model generates the PVCs images from PET images, without requiring any additional anatomical information such as MRI or CT. Our model eliminates the need for accurate registration or segmentation or PSF calculation. Also, no assumptions regarding tumor size, homogeneity, boundary, or background level are required.

• Download figure
• Open in new tab
• Download powerpoint