Deep Learning Image Reconstruction

The Deep Learning Image Reconstruction (DLIR) system employs a specifically trained Deep Neural Network to produce high-quality cross-sectional CT images for diverse clinical applications, including head, whole body, cardiac, and vascular imaging. This modified version enhances reconstruction speed and image quality on Revolution CT systems, aiming for an FBP-like image appearance while preserving or exceeding the performance of ASiR-V in noise reduction, contrast detectability, and spatial resolution.

The Deep Learning Image Reconstruction (DLIR) option is a deep learning-based method designed to produce cross-sectional CT images for head, whole body, cardiac, and vascular applications across all ages. It utilizes a dedicated Deep Neural Network (DNN) trained on the Revolution CT/Apex platform to reconstruct images from X-ray transmission data obtained via Axial, Helical (Volumetric), and Cardiac acquisitions. The core objective is to generate 'TrueFidelity CT Images' that exhibit an image appearance similar to traditional Filtered Back Projection (FBP) images while maintaining the performance levels of ASiR-V in key areas such as image noise, low contrast detectability, high-contrast spatial resolution, and streak artifact suppression. This submission pertains to a modification of the DLIR algorithm on Revolution family CT systems aimed at enhancing reconstruction speed and image quality, offering user-selectable strengths (Low, Medium, High). Non-clinical engineering bench testing demonstrated substantial equivalence by comparing the modified DLIR to ASiR-V using identical raw datasets, evaluating metrics like low contrast detectability, image noise, spatial resolution, artifact suppression, CT number uniformity and accuracy, and contrast-to-noise ratio. Specific evaluations included pediatric phantoms and low-dose lung cancer screening protocols. Clinical testing involved a reader study with 40 retrospectively collected cases, where raw data was reconstructed with both ASiR-V and the modified DLIR. Six board-certified radiologists assessed diagnostic image quality and directly compared ASiR-V and DLIR images for image noise texture, sharpness, and noise texture homogeneity. The study concluded that DLIR produces diagnostic quality images with significantly better subjective image quality than those generated with the ASiR-V reconstruction algorithm.

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. What specific architectural changes were made to the Deep Neural Network in this modified DLIR version compared to the predicate K183202, and how do these changes contribute to improved reconstruction speed and image quality?; Could you elaborate on the 'retraining and inferencing techniques' employed for the revised DNN and the nature of the dataset used for retraining?; How does the 'image appearance similar to traditional FBP images' quantitatively compare across different anatomical regions and acquisition protocols, beyond axial NPS plots?; What are the typical reconstruction time differences between Low, Medium, and High strengths of DLIR, and how do these compare to ASiR-V and traditional FBP?; Beyond the listed metrics (LCD, noise, MTF, etc.), were any specific types of artifacts (e.g., plastic-like texture sometimes seen with DL) evaluated, and how did the modified DLIR perform in those regards compared to ASiR-V?