Search Results

3DXR is the AW application which is intended to perform the three-dimensional (3D) reconstruction computation of any images acquired with a 3D Acquisition mode of the X-ray interventional system for visualization under Volume Viewer. The 3D Acquisition modes are intended for imaging vessels, bones and soft tissues as well as other internal body structures. The 3D reconstructed Volume assist the physician in diagnosis, surgical planning, Interventional procedures and treatment follow-up.

3DXR is a post-processing software-only application, runs on Advantage Workstation (AW) platform [K110834], and performs 3D reconstruction for the CBCT 3D acquisition images (input) acquired from the fixed interventional X-ray system [K181403, K232344]. The reconstructed 3D volume (output) is visualized under Volume Viewer application [K041521]. The proposed 3DXR is a modification from the predicate 3DXR [included and cleared in K181403]. A new option, called CleaRecon DL, based on Deep-Learning (DL) technology, is added in the proposed subject 3DXR application.

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

1. A table of acceptance criteria and the reported device performance

2. Sample size used for the test set and the data provenance (e.g. country of origin of the data, retrospective or prospective)

• Engineering Bench Testing:
  • Test Set 1 (Image Pairs): Two samples (number not specified beyond "two samples").
  • Test Set 2 (Phantoms): Not specified beyond "phantoms."
• Clinical Image Quality Evaluation (Retrospective):
  • Sample Size: 110 independent exams, each from a unique patient.
  • Data Provenance: Retrospectively collected from 13 clinical sites.
    • 80 patients from the US
    • 26 patients from France
    • 4 patients from Japan
  • Patient Population: Adult patients (pediatrics excluded) undergoing interventional procedures. No inclusion criteria for age (within adult range) or sex/gender (except for prostate).

3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts (e.g. radiologist with 10 years of experience)

• Engineering Bench Testing: Not applicable; ground truth was intrinsic (images with and without applying CleaRecon DL, or phantoms with reference).
• Clinical Image Quality Evaluation:
  • Number of Experts: 13 clinicians.
  • Qualifications: "Clinicians" (specific specialties or years of experience are not mentioned, but their role implies expertise in image interpretation for interventional procedures).

4. Adjudication method (e.g. 2+1, 3+1, none) for the test set

• Clinical Image Quality Evaluation: Each of the 110 exams (with and without CleaRecon DL) was compared/evaluated at least 3 times independently by the recruited clinicians. This resulted in 490 pairs of clinicians' evaluations. This suggests a multi-reader, independent review with subsequent aggregation of results, rather than a formal consensus-based adjudication like 2+1 or 3+1 for individual cases, as the "489 reviews" and "98% of cases" suggest aggregated findings.

5. 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

• A multi-reader multi-case (MRMC) study was implicitly conducted as part of the clinical image quality evaluation, with 13 clinicians reviewing 110 cases (though the "multi-case" aspect is strong, the "multi-reader" aspect is also present for each case, as each was reviewed at least 3 times independently).
• Effect Size: The study focused on the impact of the image quality on interpretation, rather than a direct measure of human reader performance improvement in diagnostic accuracy or efficiency with and without AI assistance. The results indicated:
  • "In 98% of the cases, the CBCT images reconstructed with CleaRecon DL option are evaluated as clearer than the conventional CBCT images."
  • Clinicians were asked to assess "how it impacts their confidence in image interpretation," but the specific effect size or improvement in confidence wasn't quantified.
  • No hidden or artificially created structures were identified, indicating perceived safety and reliability.
  • Therefore, while it showed a significant improvement in perceived image clarity, it did not provide a quantitative effect size for human reader improvement (e.g., in AUC, sensitivity, or specificity) with AI assistance.

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

• Yes, a standalone performance evaluation of the algorithm was done as part of the "Engineering bench testing." This involved:
  • Testing on a segregated test dataset of image pairs (with and without CleaRecon DL) where Mean Absolute Error (MAE) and Structural Similarity Index Measure (SSIM) were computed.
  • Testing on phantoms where MAE and Standard Deviation (SD) were computed relative to a reference (without simulated artifacts).
  • These tests directly assessed the algorithm's capability to reduce artifacts and improve image metrics without human interaction.

7. The type of ground truth used (expert consensus, pathology, outcomes data, etc)

• Engineering Bench Testing:
  • For the image pair comparison, the ground truth was essentially the "ideal" or "less artifact-laden" image derived from the paired comparison. This is a form of reference-based comparison where the output without the artifact or with the artifact corrected is the standard.
  • For phantoms, the ground truth was the known characteristics of the phantom (e.g., absence of artifacts in the reference image).
• Clinical Image Quality Evaluation:
  • The ground truth was established through expert evaluation/consensus (13 clinicians evaluating side-by-side images). However, it was focused on subjective image quality and the presence/absence of artifacts, rather than ground truth for a specific diagnosis or outcome.

8. The sample size for the training set

• The text states: "The CleaRecon DL algorithm was trained and qualified using pairs of images with and without streak artifacts." However, the specific sample size of the training set is not provided.

9. How the ground truth for the training set was established

• The ground truth for the training set was established using "pairs of images with and without streak artifacts." This implies that for each image with streak artifacts, there was a corresponding reference image without such artifacts, which allowed the algorithm to learn how to remove them. The method by which these "pairs of images" and their respective "with/without streak artifacts" labels were generated or confirmed is not detailed. It could involve expert labeling, simulation, or other methods.

Ask a specific question about this device