Search Results

The Deep Learning Image Reconstruction software is a deep learning based reconstruction method intended to produce cross-sectional images of the head and whole body by computer reconstruction of X-ray transmission data taken at different angles and planes, including Axial, Helical (Volumetric), and Cardiac acquisitions, for all ages. Deep Learning Image Reconstruction software can be used for head, whole body, cardiac, and vascular CT applications.

Deep Learning Image Reconstruction is an image reconstruction method that uses a dedicated Deep Neural Network (DNN) that has been designed and trained specifically to generate CT Images to give an image appearance, as shown on axial NPS plots, similar to traditional FBP images while maintaining the performance of ASiR-V in the following areas: dose, image noise (pixel standard deviation), low contrast detectability, high-contrast spatial resolution, and streak artifact suppression.

The images produced are branded as "TrueFidelity"" CT Images". Reconstruction times with Deep Learning Image Reconstruction software support a normal throughput for routine CT.

The deep learning technology is integrated into the scanner's existing raw data-based image reconstruction chain to produce DICOM compatible "TrueFidelity"™ CT Images".

The system allows user selection of three strengths of Deep Learning Image Recon: Low, Medium or High. The strength selection will vary with individual users' preferences and experience for the specific clinical need.

Here's a breakdown of the acceptance criteria and the study that proves the device meets them, based on the provided text:

Acceptance Criteria and Device Performance

The core of the acceptance criteria revolves around demonstrating that the Deep Learning Image Reconstruction (DLIR) on the Revolution Ascend system is substantially equivalent to its predicate device (DLIR on Revolution EVO) and performs at least as well as, or better than, ASiR-V reconstruction in key image quality metrics.

Study Details

1. Sample size used for the test set and the data provenance:

   • Sample Size: A total of 60 retrospectively collected clinical cases were used.
   • Data Provenance: The data was retrospectively collected. The country of origin is not explicitly stated in the provided text.

2. Number of experts used to establish the ground truth for the test set and the qualifications of those experts:

   • Number of Experts: 9 board-certified radiologists.
   • Qualifications: These radiologists had "expertise in the specialty areas that align with the anatomical region of each case."

3. Adjudication method for the test set:

   • Each image was read by 3 different radiologists.
   • The readers completed their evaluations independently and were blinded to the results of the other readers' assessments.
   • The text doesn't explicitly state an adjudication method like 2+1 or 3+1 for discrepancies. It implies a consensus or agreement was sought, or that individual assessments contributed to the overall conclusion of substantial equivalence. Given they provided an assessment on a Likert scale and then compared images, it seems individual reader assessments were aggregated, rather than a discrepancy resolution process.

4. If a multi reader multi case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance:

   • Yes, an MRMC study was implicitly done, as 9 radiologists evaluated 60 cases, with each case being read by 3 different radiologists. The study involved a comparison between ASiR-V reconstructions and Deep Learning Image Reconstruction (DLIR) images.
   • Effect Size: The document does not provide a specific effect size (e.g., percentage improvement in accuracy or AUC) of how much human readers improved with DLIR assistance compared to ASiR-V. It states that the study results "support substantial equivalence and performance claims" and that readers assessed image quality and compared noise texture and sharpness, implying favorable or equivalent performance.

5. If a standalone (i.e. algorithm only without human-in-the-loop performance) was done:

   • Yes, extensive standalone (algorithm only) non-clinical engineering bench testing was performed. This included evaluations of:
     • Low contrast detectability (LCD)
     • Image Noise (pixel standard deviation)
     • High contrast spatial resolution (MTF)
     • Streak Artifact Suppression
     • Spatial Resolution
     • Noise Power Spectrum (NPS) and Standard Deviation of noise
     • CT Number Accuracy and Uniformity
     • Contrast to Noise (CNR) ratio
     • Artifact analysis - metal objects, unintended motion, truncation
     • Pediatric Image Quality Performance
     • Low Dose Lung Cancer Screening

6. The type of ground truth used:

   • For the clinical reader study, the ground truth was based on expert assessment/consensus (implying the "gold standard" for diagnostic image quality, noise texture, and sharpness was the radiologists' expert opinion). The cases were "retrospectively collected clinical cases," suggesting the presence of a known clinical diagnosis or outcome, but the specific ground truth for disease presence/absence is not explicitly stated as the primary output of the DLIR evaluation. The evaluation focused more on image quality attributes and comparison between reconstruction methods rather than diagnostic accuracy against a separate definitive truth.

7. The sample size for the training set:

   • The document states the Deep Neural Network (DNN) was "trained on the Revolution family CT Scanners" but does not provide the specific sample size (number of images or cases) used for training.

8. How the ground truth for the training set was established:

   • The text does not explicitly detail how the ground truth for the training set was established. It mentions the DNN was "designed and trained specifically to generate CT Images to give an image appearance... similar to traditional FBP images while maintaining the performance of ASiR-V in the following areas: dose, image noise (pixel standard deviation), low contrast detectability, high-contrast spatial resolution, and streak artifact suppression." This suggests the training likely involved pairing raw CT data with expertly reconstructed ASiR-V or FBP images as a reference for image quality characteristics. The ground truth in this context would be the desired output image characteristics (e.g., low noise, high resolution) that the DLIR algorithm was optimized to reproduce.

Ask a specific question about this device