HYPER DLR

HYPER DLR is a software-only medical device intended to enhance fluorodeoxyglucose (FDG) PET image quality by reducing noise through a pre-trained convolutional neural network, functioning as an image post-processing technique. This method aims to offer superior noise reduction and image detail preservation compared to traditional Gaussian filtering, thereby potentially improving diagnostic confidence for clinicians using compatible uMI 550 and uMI 780 PET/CT systems.

HYPER DLR is a software-only device designed for image post-processing on existing uMI 550 and uMI 780 PET/CT systems. Its primary purpose is to reduce noise in fluorodeoxyglucose (FDG) PET images, thereby aiding radiologists and nuclear medicine physicians in their diagnostic interpretations. The core technology involves a pre-trained convolutional neural network that processes high-noise PET images to predict and generate low-noise versions by specifically extracting the noise component, distinguishing it from useful image details, which offers an alternative to traditional image smoothing techniques like Gaussian filtering. The performance of HYPER DLR was assessed through non-clinical testing, including engineering bench tests and clinical image evaluations, without conducting new clinical studies. Bench testing compared HYPER DLR against Gaussian filtering using identical raw datasets from uMI 780 and uMI 550 systems, evaluating metrics such as peak signal-to-noise ratio, structural similarity index, SNR, and CNR. These tests demonstrated that HYPER DLR significantly improved SNR and CNR while preserving image consistency. Clinical image evaluations by board-certified nuclear medicine physicians indicated that HYPER DLR produced images with lower noise than Gaussian filtering and maintained sufficient diagnostic quality, an observation consistent across typical clinical scan times for the compatible PET/CT systems.

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 DLR vary with different levels of initial image noise and varying radiotracer uptake in FDG PET scans beyond the tested scenarios?; What are the specific architectural details and training dataset characteristics of the pre-trained convolutional neural network in HYPER DLR that ensure its robustness and generalizability across diverse patient demographics and acquisition protocols?; Beyond the reported qualitative improvements and standard image quality metrics (SNR, CNR), how does HYPER DLR quantitatively affect standardized uptake values (SUVs) and lesion detectability, particularly for small or low-contrast lesions, compared to Gaussian filtering?; Are there any specific types of image artifacts that HYPER DLR might introduce or be less effective at mitigating compared to traditional methods, and what is the learning curve for physicians to interpret HYPER DLR-processed images?; What is the computational overhead (processing time per study) of HYPER DLR when integrated into the existing workflow of uMI 550 and uMI 780 systems, and how does it compare to the processing time of Gaussian filtering?