Deep Learning Image Reconstruction

The Deep Learning Image Reconstruction software employs a dedicated Deep Neural Network for generating CT images across various anatomical regions and patient populations, aiming to replicate traditional image appearance while enhancing quality metrics like noise reduction and spatial resolution. This iteration ports the established algorithm to the Discovery CT750 HD family of CT systems, leveraging GE's Edison Platform for deployment.

GE Healthcare's Deep Learning Image Reconstruction (DLIR) is a software designed to produce cross-sectional computed tomography (CT) images of the head and whole body. It utilizes a dedicated Deep Neural Network (DNN) to reconstruct images from X-ray transmission data acquired through Axial, Helical (Volumetric), and Cardiac scans, suitable for all patient ages. The software is intended for head, whole body, cardiac, and vascular CT applications, producing images branded as "TrueFidelityTM CT Images" and offering user-selectable strengths of Low, Medium, or High for image reconstruction. The DLIR software's DNN is specifically trained to generate CT images with an appearance comparable to traditional Filtered Back Projection (FBP) methods, while aiming to achieve performance levels equal to or better than GE's ASiR-V technology in terms of radiation dose efficiency, image noise (pixel standard deviation), low contrast detectability, high-contrast spatial resolution, and streak artifact suppression. This submission details the porting of the existing DLIR algorithm to the Discovery CT750 HD family of CT systems (including Discovery CT750 HD, Revolution Frontier, and Revolution Discovery CT). While integrated into the scanner's raw data processing chain, the deployment environment for this version is GE's Edison Platform, a change from the predicate's on-console deployment. Extensive software and IQ bench testing confirmed that the DLIR software meets its design requirements and performance criteria on these systems, maintaining image quality and diagnostic performance.

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 TrueFidelity CT Images generated by this DLIR software on the Discovery CT750 HD family compare quantitatively to traditional FBP and ASiR-V across different anatomical regions, patient demographics, and varying radiation dose levels?; What are the specific architectural characteristics of the Deep Neural Network employed, and how was its training dataset curated and validated to ensure robust performance and generalizability across diverse clinical cases and acquisition protocols on the Discovery CT750 HD systems?; Are there any known limitations or specific clinical scenarios where the different strengths (Low, Medium, High) of the DLIR software might introduce subtle image artifacts or misrepresent anatomical details compared to ASiR-V or FBP, particularly in challenging cases like pediatric or bariatric imaging?; How does the selection of DLIR strength (Low, Medium, High) quantitatively impact image noise, low contrast detectability, spatial resolution, and Hounsfield Unit accuracy, and what is the corresponding potential for radiation dose reduction while maintaining diagnostic image quality?; What is the computational demand and typical image reconstruction time for DLIR on the Discovery CT750 HD systems when deployed via GE's Edison Platform, and how does this compare to the predicate device's on-console deployment and other reconstruction methods like ASiR-V, considering clinical workflow efficiency?