Reston, VA – A new method of artificial intelligence can be used to generate high-quality “PET/CT” images and subsequently reduce patient exposure to radiation. Developed by the National Cancer Institute, the method circumvents the need for computed tomography-based attenuation correction, potentially allowing more frequent PET imaging to monitor disease and treatment progression without radiation exposure from the acquisition of computed tomography. This research was presented at the 2022 Annual Meeting of the Society of Nuclear Medicine and Molecular Imaging.
Cancer patients often undergo multiple imaging studies throughout diagnosis and treatment, potentially including multiple PET/CT scans in close succession. The CT portion of the exam contributes to a patient’s overall radiation exposure, but is largely redundant. In this study, researchers sought to reduce or eliminate the need for low-dose PET/CT scans by using an artificial intelligence model to generate attenuation-corrected virtual PET scans.
The data cohort for the development of artificial intelligence models included 305 18F-DCFPyL PSMA PET/CT studies. Each study contained three scans: non-attenuation-corrected PET, attenuation-corrected PET, and low-dose CT. The studies were divided into three sets for training (185), validation (60), and testing (60). A 2D Pix2Pix generator was then used to generate synthetic attenuation-corrected PET (gen-PET) scans from the original non-attenuation-corrected PET.
For the qualitative assessment, two nuclear medicine physicians reviewed 40 PET/CT studies in random order, without knowing whether the image was from an original attenuation-corrected PET or from a PET-gen. Each expert recorded the number and location of PET-positive lesions and qualitatively examined overall noise and image quality. Readers were able to successfully detect lesions on gen-PET images with reasonable sensitivity values.
“High-quality artificial intelligence-generated images preserve vital information from raw PET images without the added radiation exposure of CT scans,” said Kevin Ma, PhD, postdoctoral researcher at the National Cancer Institute in Bethesda, Maryland. “This opens up opportunities to increase the frequency and number of PET scans per patient per year, which could provide more accurate assessment of lesion detection, treatment efficacy, radiotracer efficacy and other measures in research and patient care.”
|Figure 1. Representative axial image of a test set imaging study. From top left to bottom right: CT (A), NAC-PET (B), original AC-PET (C), AC-PET generated by AI (D). The blue arrow indicates a lesion that was observed in both images, and the red arrow indicates a lesion that was missed (i.e., not detected by nuclear medicine physicians) in the AI image. NAC = uncorrected for attenuation. AC = Corrected Attenuation.|
Abstract 151. “Artificial Intelligence-Generated PET Images for PSMA-PET/CT Studies: Quantitative and Qualitative Assessment,” Kevin Ma, National Cancer Institute, National Institutes of Health, College Park, Maryland; Esther Mena, Liza Lindenberg, Deborah Citrin, William Dahut, James Gulley, Peter Choyke, Baris Turkbey, and Stephanie Harmon, National Cancer Institute, National Institutes of Health, Bethesda, Maryland; Peter Pinto, Urologic Oncology Branch, National Cancer Institute, National Institutes of Health, Bethesda, Maryland; Bradford Wood, Radiology and Imaging Sciences, Cancer Research Center, National Cancer Institute, National Institutes of Health, Bethesda, Maryland; and Ravi Madan, Genitourinary Malignant Tumors Branch, National Cancer Institute, National Institutes of Health, Bethesda, Maryland.