Deep Learning Image Reconstruction

The Deep Learning Image Reconstruction (DLIR) software employs a pre-trained Deep Neural Network to enhance computed tomography image quality by reconstructing cross-sectional images with an appearance similar to Filtered Back Projection but with the performance benefits of ASiR-V. This software iteration validates the porting of this existing DLIR algorithm to the Revolution Ascend CT system, demonstrating comparable or superior image noise reduction, low contrast detectability, and spatial resolution through rigorous non-clinical and clinical evaluations.

The Deep Learning Image Reconstruction (DLIR) software by GE Healthcare is a sophisticated image processing tool designed for computed tomography (CT) systems. Its primary purpose is to reconstruct cross-sectional images of the head and whole body, accommodating various acquisition modes like Axial, Helical (Volumetric), and Cardiac scans, and is applicable to patients of all ages for head, whole body, cardiac, and vascular CT applications. The core technology relies on a dedicated Deep Neural Network (DNN) specifically trained to generate CT images, branded as "TrueFidelityTM CT Images." This DNN aims to produce images with an appearance comparable to traditional Filtered Back Projection (FBP) methods while maintaining or exceeding the performance benchmarks of GE's existing ASiR-V reconstruction technique in critical areas such as radiation dose efficiency, image noise reduction, low-contrast object detectability, high-contrast spatial resolution, and suppression of streak artifacts. The DLIR software is integrated into the CT scanner's raw data reconstruction chain and offers users selectable strengths (Low, Medium, High) to tailor image quality. This particular application details the porting of the established DLIR algorithm, without retraining the neural network, to the Revolution Ascend CT system, extending its availability from previous Revolution CT and EVO platforms. To validate its performance on the Revolution Ascend, GE Healthcare conducted comprehensive non-clinical and clinical evaluations. Non-clinical bench testing involved comparing DLIR-reconstructed images with those from ASiR-V using identical raw datasets from the Revolution Ascend. This testing assessed numerous image quality parameters, including low contrast detectability, image noise, spatial resolution, artifact suppression, CT number accuracy, and performance in pediatric and low-dose lung cancer screening scenarios, confirming performance as good or better than ASiR-V. A clinical reader study involving 60 retrospective cases, reviewed by 9 board-certified radiologists, further supported these findings, with radiologists assessing diagnostic image quality and directly comparing DLIR images to ASiR-V images on noise texture and sharpness, affirming the software's effectiveness for its intended use on the Revolution Ascend system.

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 the DLIR algorithm, specifically regarding low-contrast detectability and noise texture, vary across the three user-selectable strengths (Low, Medium, High) on the Revolution Ascend system, and are there established guidelines for selecting the optimal strength for different anatomical regions or clinical questions?; Given that the Deep Neural Network was not retrained for the Revolution Ascend system, what specific non-clinical tests were performed to ensure its robustness and generalizability across different hardware (Revolution EVO vs. Revolution Ascend) and potential variations in raw data characteristics?; The summary mentions pediatric image quality performance testing; could specific quantitative or qualitative improvements be detailed for pediatric protocols when using DLIR compared to ASiR-V, particularly concerning dose reduction potential while maintaining diagnostic quality?; How does the reconstruction time of DLIR on the Revolution Ascend system compare to ASiR-V and traditional Filtered Back Projection, and does this impact clinical workflow efficiency, especially in high-throughput environments or emergency settings?; Were there any specific types of artifacts or challenging clinical scenarios (e.g., presence of dense metallic implants, significant patient motion) where DLIR demonstrated a more notable improvement or limitation compared to ASiR-V in the clinical reader study or bench testing?