Search Results

Sonic DL is a Deep Learning based image reconstruction technique that is available for use on GE Healthcare 1.5T and 3.0T MR systems. Sonic DL reconstructs MR images from highly under-sampled data, and thereby enables highly accelerated acquisitions. Sonic DL is intended for cardiac imaging, and for patients of all ages.

Sonic DL is a new software feature intended for use with GE Healthcare MR systems. It consists of a deep learning based reconstruction algorithm that is applied to data from MR cardiac cine exams obtained using a highly accelerated acquisition technique.

Sonic DL is an optional feature that is integrated into the MR system software and activated through a purchasable software option key.

Here's a breakdown of the acceptance criteria and the study details for the Sonic DL device, based on the provided document:

Sonic DL Acceptance Criteria and Study Details

1. Table of Acceptance Criteria and Reported Device Performance

The document describes the performance of Sonic DL in comparison to conventional ASSET Cine images. While explicit numerical acceptance criteria for regulatory clearance are not stated, the studies aim to demonstrate non-inferiority or superiority in certain aspects. The implicit acceptance criteria are:

• Diagnostic Quality: Sonic DL images must be rated as being of diagnostic quality.
• Functional Measurement Agreement: Functional cardiac measurements (e.g., LV volumes, EF, CO) from Sonic DL images must agree closely with those from conventional ASSET Cine images, ideally within typical inter-reader variability.
• Reduced Scan Time: Sonic DL must provide significantly shorter scan times.
• Preserved Image Quality: Image quality must be preserved despite higher acceleration.
• Single Heartbeat Imaging (Functional): Enable functional imaging in a single heartbeat.
• Rapid Free-Breathing Functional Imaging: Enable rapid functional imaging without breath-holds.

2. Sample Size Used for the Test Set and Data Provenance

The document describes two primary reader evaluation studies and additional clinical testing.

• First Reader Study (Functional Measurements):
  • Sample Size: 107 image series from 57 unique subjects (46 patients, 11 healthy volunteers).
  • Data Provenance: Data from 7 sites: 2 GE Healthcare facilities and 5 external clinical collaborators. This indicates data from multiple sources, likely a mix of prospective and retrospective collection. The geographic origin of these sites is not explicitly stated but implies a multi-center study potentially from different countries where GE Healthcare operates or collaborates.
• Second Reader Study (Image Quality Assessment):
  • Sample Size: 127 image sets, which included a subset of the subjects from the first study.
  • Data Provenance: Same as the first reader study (clinical sites and healthy volunteers at GE Healthcare facilities).
• Additional Clinical Testing (1 R-R Free Breathing):
  • Functional Measurements: 10 subjects.
  • Image Quality Evaluation: 13 subjects.
  • Data Provenance: In vivo cardiac cine images from 19 healthy volunteers. This implies prospective collection or a subset of prospectively collected healthy volunteer data.

3. Number of Experts Used to Establish the Ground Truth for the Test Set and Qualifications

• First Reader Study (Functional Measurements): Three radiologists. Qualifications are not explicitly stated, but their role in making quantitative measurements implies expertise in cardiac MRI.
• Second Reader Study (Image Quality Assessment): Three radiologists. Qualifications are not explicitly stated, but their role in blinded image quality assessments implies expertise in cardiac MRI interpretation.
• Additional Clinical Testing (1 R-R Free Breathing Image Quality): One U.S. board certified radiologist.

4. Adjudication Method for the Test Set

The document does not explicitly state an adjudication method (like 2+1, 3+1, or none) for either the functional measurements or the image quality assessments. For the first study, it mentions "inter-method variability" and "inter-observer intra-method variability," suggesting that the readings from the three radiologists were compared against each other and against the conventional method, but not necessarily adjudicated to establish a single "ground truth" per case. For the second study, "blinded image quality assessments" were performed, and ratings were averaged, but no adjudication process is described.

5. If a Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study was done, and the effect size

A clear MRMC comparative effectiveness study, in the sense of measuring human reader improvement with AI vs. without AI assistance, is not explicitly described.

The studies compare the performance of Sonic DL images (algorithm output) against conventional images, with human readers evaluating both.

• The first reader study compares quantitative measurements from Sonic DL images to conventional images, indicating suitability for performing functional cardiac measurements by showing smaller inter-method variability than inter-observer intra-method variability for conventional images. This suggests Sonic DL is at least as reliable as the variability between conventional human measurements.
• The second reader study involves blinded image quality assessments of both conventional and Sonic DL images, confirming that Sonic DL images were rated as diagnostic quality.
• The additional clinical testing for 1 R-R free breathing shows that Sonic DL images were "scored higher than the corresponding conventional free breathing Cine images" by a U.S. board-certified radiologist.

These are comparisons of the image quality and output from the AI system versus conventional imaging, interpreted by readers, rather than measuring human reader performance assisted by the AI system.

Therefore, the effect size of how much human readers improve with AI vs. without AI assistance is not provided because the studies were designed to evaluate the image output quality and measurement agreement of the AI-reconstructed images themselves, not to assess an AI-assisted workflow for human readers.

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

Yes, standalone performance was assessed for image quality metrics.

• Nonclinical Testing: "Model accuracy metrics such as Peak-Signal-to-Noise (PSNR), Root-Mean-Square Error (RMSE), Structural Similarity Index Measure (SSIM), and Mean Absolute Error (MAE) were used to compare simulated Sonic DL images with different levels of acceleration and numbers of phases to the fully sampled images." This is a standalone evaluation of the algorithm's output quality against a reference.
• In Vivo Testing: "model accuracy and temporal sharpness evaluations were conducted using in vivo cardiac cine images obtained from 19 health volunteers." This is also a standalone technical evaluation of the algorithm's output on real data.

7. The Type of Ground Truth Used

• Nonclinical Testing (Simulated Data): The ground truth was the "fully sampled images" generated from an MRXCAT phantom and a digital phantom.
• Clinical Testing (Reader Studies):
  • Functional Measurements: The "ground truth" for comparison was the measurements taken from the "conventional ASSET Cine images." The variability of these conventional measurements across readers also served as a baseline for comparison. This is a form of clinical surrogate ground truth (comparing to an established accepted method).
  • Image Quality Assessments: The "ground truth" was the expert consensus/opinion of the radiologists during their blinded assessments of diagnostic quality.
  • Additional Clinical Testing (1 R-R Free Breathing): Functional measurements were compared to "conventional ASSET Cine breath hold images" (clinical surrogate ground truth). Image quality was based on the scoring by a "U.S. board certified radiologist" (expert opinion).

No pathology or outcomes data were used as ground truth. The ground truth in the clinical setting was primarily based on established imaging techniques (conventional MR) and expert radiologist assessments.

8. The Sample Size for the Training Set

The document does not explicitly state the sample size for the training set used for the deep learning model. It only describes the data used for testing the device.

9. How the Ground Truth for the Training Set Was Established

Since the training set size is not provided, the method for establishing its ground truth is also not described in the provided text. Typically, for deep learning reconstruction, the "ground truth" for training often involves fully sampled or high-quality reference images corresponding to the undersampled input data.

Ask a specific question about this device

AIR Recon DL is a deep learning based reconstruction technique that is available for use on GE Healthcare 1.5T, 3.0T, and 7.0T MR systems. AIR Recon DL reduces noise and ringing (truncation artifacts) in MR images, which can be used to reduce scan time and improve image quality. AIR Recon DL is intended for use with all anatomies, and for patients of all ages. Depending on the anatomy of interest being imaged, contrast agents may be used.

AIR Recon DL is a software feature intended for use with GE Healthcare MR systems. It is a deep learning based reconstruction technique that removes noise and ringing (truncation) artifacts from MR images. AIR Recon DL is an optional feature that is integrated into the MR system software and activated through a purchasable software option key.

Here's a breakdown of the acceptance criteria and the study that proves the device meets them, based on the provided FDA 510(k) summary for AIR Recon DL:

1. Acceptance Criteria and Reported Device Performance

The acceptance criteria are implicitly defined by the desired improvements and non-inferiority claims of the AIR Recon DL feature. The reported device performance demonstrates how these criteria were met.

2. Sample Size and Data Provenance

• Test Set Sample Size: 133 cases.
  • 129 patient cases
  • 4 healthy subjects
• Data Provenance:
  • Country of Origin: Not explicitly stated, but "10 different clinical sites" suggests a multi-center study, and "a GE Healthcare facility" could indicate a US or international site. Given the FDA submission, it's highly likely to include US data.
  • Retrospective or Prospective: Not explicitly stated, but the description "images acquired across a variety of pulse sequences and anatomies" and involvement of "10 different clinical sites" could imply a retrospective collection for the reader study, where images were pre-collected. However, without explicit mention, it's not definitive. The phrasing "acquired from the same acquired raw data" suggests a paired comparison based on existing data.

3. Number of Experts and Qualifications

• Number of Experts: Three radiologists.
• Qualifications: "Radiologists" implies medical doctors specialized in radiology. No further specifics on years of experience or subspecialty were provided.

4. Adjudication Method for the Test Set

The adjudication method appears to be 2+1 (or 3/3 agreement is ideal, but 2 out of 3 for consensus is common).
"Each image pair was evaluated independently by three radiologists."
"The results confirmed that the AIR Recon DL feature provides images with equivalent or better image quality in terms of apparent signal to noise ratio (133 out of 133 cases), sharpness (133 out of 133 cases), and lesion conspicuity (123 out of 124 cases with pathology)."
"The radiologists reading the images also indicated a preference for the AIR Recon DL images over conventional images in 99% of the evaluations."
This indicates that the claims are based on the collective agreement or majority opinion of the three readers for each case. The exact decision rule (e.g., simple majority, unanimous) is not stated, but the high consistency (e.g., 133/133, 123/124) implies strong agreement or effective resolution.

5. Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study

Yes, a multi-reader multi-case (MRMC) comparative effectiveness study was done.

• Comparison: Radiologists compared AIR Recon DL images to conventional images (without AIR Recon DL) reconstructed from the same acquired raw data.
• Effect Size of Human Reader Improvement: The study demonstrates a significant preference and improvement in perceived image quality attributes by human readers when using AIR Recon DL assisted images compared to conventional images.
  • Preference: Radiologists preferred AIR Recon DL images over conventional images in 99% of evaluations. This indicates a strong positive effect on reader perception and diagnostic confidence.
  • Qualitative Improvement:
    • Equivalent or better SNR in 100% of cases (133/133).
    • Equivalent or better sharpness in 100% of cases (133/133).
    • Equivalent or better lesion conspicuity in ~99.2% of pathological cases (123/124).
  • Enablement of Shorter Scans: For shorter scan time acquisitions, AIR Recon DL images were rated as better or equivalent image quality in 100% of 22 image pairs, which suggests human readers are able to maintain or even improve their assessment quality despite reduced acquisition time. Overall, the effect size is very large and consistently positive across all measured subjective criteria.

6. Standalone (Algorithm Only) Performance

The document describes "nonclinical testing" on phantoms, which represents a form of standalone testing where the algorithm's output is directly measured against predefined physical criteria:

• "AIR Recon DL has undergone phantom testing to evaluate the feature and its impact on image quality, including SNR, sharpness, and low contrast detectability."
• "The nonclinical testing demonstrated that AIR Recon DL does improve SNR and image sharpness while maintaining low contrast detectability."
• "ADC maps were not adversely impacted by the use of AIR Recon DL."
  This evaluates the algorithm's effect on image characteristics absent human interpretation of clinical cases.

7. Type of Ground Truth Used

The study primarily used expert consensus (radiologist agreement) as the ground truth for evaluating image quality attributes (SNR, sharpness, lesion conspicuity, overall preference) and the impact of the algorithm.
For the "presence of pathology" in the test set, it's assumed that this was either identified beforehand from clinical reports or pathology (e.g., biopsy results) or by consensus among the evaluating radiologists prior to their evaluation of the AI-enhanced images. However, the exact source of ground truth for pathology presence/absence isn't detailed, only that 124 cases "with pathology" and 9 cases "without pathology" were included.

8. Sample Size for the Training Set

The document states: "Both the proposed AIR Recon DL and the predicate device use neural networks that have similar architecture, and were trained using similar methods and data."
However, the specific sample size for the training set is NOT provided in this summary.

9. How Ground Truth for Training Set was Established

The document states: "Both the proposed AIR Recon DL and the predicate device use neural networks that have similar architecture, and were trained using similar methods and data."
The method for establishing ground truth for the training set is NOT explicitly detailed. Typically, for deep learning reconstruction, the "ground truth" during training often refers to high-quality, fully sampled MR images (or images from a prior, high-quality reconstruction method) that the AI attempts to match or improve upon, rather than a clinical diagnosis per se. The goal during training would be to generate images that are less noisy and sharper while preserving underlying anatomical and pathological information, learned by comparing "corrupted" (e.g., undersampled, noisy) inputs to "clean" reference images.

Ask a specific question about this device