AI-assisted pre-scan PET MIP guide the position of slab in PET/MR loco-regional imaging

Introduction: Compare with PET/CT, PET/MR appears more promising for loco-regional staging of diseases for which MR is the anatomical imaging modality of choice [1]. MR imaging is usually a scan of the required body part, however, for a PET/MR loco-regional imaging, the cover of primary tumor and regional lymph node metastasis is necessary. For MR images are less sensitive to lesions than PET images, position of slab assisted on PET images may be a better choice in PET/MR loco-regional imaging. Before the loco-regional imaging, PET/MR or PET/CT whole-body images should be acquired in advance to get the distribution of lesions, but the whole-body PET image may can’t be used to guide slab positioning directly for the subsequent loco-regional imaging because of different Patient Coordinate System (PCS). The causes of different PCS include different scanner, resetting the PCS through laser alignment and the change of patient position. We propose an AI-assisted pre-scan PET MIP (termed AI pre-PET) here to work as a substitution of whole-body PET to guide the slab positioning. For whole-body pre-PET imaging last less than 60 seconds and provides the distribution of focal lesions, it may increase the efficiency of PET/MR loco-regional acquisition by optimize the slab positioning workflow.

Methods: The cohort of this post hoc analysis consisted of 110 patients, including 24 ⁶⁸Ga-PSMA PET and 86 ¹⁸F-FDG PET. All patients underwent PET/MR examination using a United Imaging uPMR790 scanner (United Imaging Healthcare Co. Ltd., Shanghai, China). List-mode PET imaging was performed at a median uptake time of 60 minutes. The standard imaging typically lasted 40 minutes, eight minutes per bed position (bp). Just before the routine PET/MR imaging, a 60 seconds pre-PET scan (12s/bp) protocol was started simultaneously with MR five-bed position whole-body localizer scan. PET and pre-PET were reconstructed using an Ordered Subsets Expectation Maximization iterative reconstruction algorithm (OSEM) with two iterations, 20 subsets and matrix size 150×150. Images were spatially smoothed with 3 mm full-width at half-maximum (FWHM) Gaussian kernels. PET emission data were corrected for scatter, random, dead time, and attenuation. A 7-layer U-Net Assisted Auto-encoder was used to generate AI-assisted pre-scan PET MIP (Fig. 1A). In order to obtain the de-noising and high-resolution training images, the enhanced PET images was generated by a pixel2pixel Generative Adversarial Network (GAN) which taken pre-PET as input, standard whole-body 40 minutes PET image worked as the gold standard of the classifier of the GAN. The training process was finished after 1000 epochs. Two experienced nuclear medicine physicians worked together to assess the image quality (five-point scale) and identify whether lesion could be detected by pre-PET and AI pre-PET compared with standard PET MIP.

Results: The pre-PET is a noisy PET image without MR-based attenuation correction (MRAC), it highlight intense uptake lesion and got an image quality score 1.5±0.71. The AI pre-PET presented an image contrast similar to that image with MRAC, compare with pre-PET, it has a better signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) and got an image quality score 2.5±0.71. Standard PET image got an image quality score 4.5±0.71. For sensitivity and specificity analysis in intense uptake lesion detection, AI pre-PET is the same as pre-PET (63.6%, 94.5%).

Conclusions: AI-assisted pre-scan PET MIP can highlight intense uptake lesions and can work as a substitution of whole-body PET to guide the slab positioning because of the high specificity. But AI pre-PET is not a substitution for standard whole-body PET scan because of the low sensitivity.

• Download figure
• Open in new tab
• Download powerpoint