HYPER AiR

HYPER AiR is a software-based image processing function designed to improve the quality of fluorodeoxyglucose (FDG) PET images by reducing noise and enhancing contrast. It achieves this by innovatively incorporating pre-trained convolutional neural networks into the established Ordered Subsets Expectation Maximization (OSEM) iterative reconstruction process, leveraging AI to distinguish and suppress noise while preserving image details.

HYPER AiR is a software-only image reconstruction technique designed for Emission Computed Tomography (ECT) systems, specifically for fluorodeoxyglucose (FDG) Positron Emission Tomography (PET) images. Its primary purpose is to enhance image quality for radiologists and nuclear medicine physicians by reducing noise and improving contrast. This software is intended for implementation on existing PET/CT devices, namely the uMI 550 and uMI 780, serving as an advanced alternative to their standard image reconstruction algorithms. The core science of HYPER AiR involves integrating pre-trained convolutional neural networks (CNNs) into the iterative Ordered Subsets Expectation Maximization (OSEM) reconstruction process. Unlike conventional OSEM, which relies on maximum-likelihood estimation techniques, HYPER AiR's CNN component is designed to distinguish between image noise and essential image details, selectively removing the noise and enhancing details. Performance verification using raw datasets from uMI 780 and uMI 550 scanners demonstrated HYPER AiR's ability to improve image contrast while suppressing background noise. Clinical image evaluations by three board-certified nuclear medicine physicians further confirmed that HYPER AiR produces images with better contrast and lower noise compared to traditional OSEM, with image quality deemed sufficient for clinical diagnosis.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the performance of HYPER AiR's pre-trained neural networks vary with different patient populations, FDG uptake levels, or acquisition protocols beyond those tested?; What specific architectures or training methodologies were used for the convolutional neural networks in HYPER AiR, and how were they optimized to distinguish noise from true image detail in FDG PET scans?; Are there any known limitations or potential artifacts introduced by the HYPER AiR algorithm, particularly in low-count statistics or unusual tracer distributions, compared to traditional OSEM?; How does the processing time for HYPER AiR reconstruction compare to standard OSEM on the uMI 550 and uMI 780 systems, and does this impact clinical workflow?; Could the HYPER AiR framework be adapted or retrained for use with other PET tracers beyond FDG, or for other emission tomography modalities?