Search Results

The Deep Learning Image Reconstruction option 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 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: 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 support a normal throughput for routine CT.

Deep Learning Image Reconstruction was trained specifically on the Revolution CT/Apex platform (K163213, K133705, K19177). 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.

The DLR algorithm is now being modified on the Revolution family CT systems (K133705, K163213, K19177) for improved reconstruction speed and image quality, thus triggering this premarket notification.

The provided document describes the Deep Learning Image Reconstruction (DLIR) device, its acceptance criteria, and the study conducted to prove it meets these criteria.

Here's a breakdown of the requested information:

1. Table of Acceptance Criteria and Reported Device Performance

The document does not explicitly state numerical acceptance criteria with pass/fail thresholds for each metric. Instead, it focuses on demonstrating non-inferiority or improvement compared to a predicate device (ASiR-V) and ensuring diagnostic quality. The reported device performance is qualitative, indicating "significantly better subjective image quality" and "diagnostic quality images."

However, based on the non-clinical and clinical testing sections, we can infer the performance metrics evaluated.

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

• Sample Size: 40 retrospectively collected clinical cases.
• Data Provenance: Retrospectively collected clinical cases. The country of origin is not specified in the provided text.

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

• Number of Experts: 6 board-certified radiologists.
• Qualifications of Experts: Board-certified radiologists with "expecialty areas that align with the anatomical region of each case."

4. Adjudication Method for the Test Set

The document describes a reader study where each of the 40 cases (reconstructed with both ASiR-V and DLIR) was read by 3 different radiologists independently. They provided an assessment of image quality using a 5-point Likert scale. There's no explicit mention of an adjudication process (e.g., 2+1, 3+1) if there were disagreements among the three readers, as the focus seems to be on independent assessment and overall subjective preference comparison.

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

Yes, a multi-reader multi-case (MRMC) comparative effectiveness study was done. Human readers compared images reconstructed with DLIR (AI-assisted reconstruction) against images reconstructed with ASiR-V (without DLIR).

• Effect Size: The study confirmed that DLIR (the subject device) produced diagnostic quality images and "have significantly better subjective image quality" than the corresponding images generated with the ASiR-V reconstruction algorithm. The text doesn't provide a specific numerical effect size (e.g., a specific improvement percentage or statistical metric), but it qualitatively states a "significant" improvement based on reader preference for image noise texture, image sharpness, and image noise texture homogeneity.

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

Yes, extensive standalone (algorithm-only) performance testing was conducted. This is detailed in the "Additional Non-Clinical Testing" section, where DLIR and ASiR-V reconstructions of identical raw datasets were compared for various objective image quality metrics without human interpretation during these specific tests.

7. The Type of Ground Truth Used

The ground truth for the clinical reader study was established through expert consensus/assessment of image quality and preference by the participating radiologists. For the non-clinical bench testing, the ground truth was based on objective physical measurements and established phantom data with known properties.

8. The Sample Size for the Training Set

The document mentions that the Deep Neural Network (DNN) for DLIR was "trained specifically on the Revolution CT/Apex platform." However, it does not specify the sample size (number of images or cases) used for the training set.

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

The text states that the DNN was "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." It also notes that the DNN "models the scanned object using information obtained from extensive phantom and clinical data."

While the exact method for establishing ground truth for training isn't explicitly detailed, it implies a process where:

• Reference Images: Traditional FBP (Filtered Back Projection) and ASiR-V images likely served as reference or target outputs for the DNN, specifically regarding image appearance, noise characteristics, and spatial resolution.
• "Extensive phantom and clinical data": This data, likely corresponding to various anatomical regions, pathologies, and dose levels, was fed into the training process. The ground truth would involve teaching the network to reconstruct images that, when compared to conventionally reconstructed images (FBP/ASiR-V), exhibit desired image quality attributes (e.g., reduced noise while preserving detail).
• Noise Modeling: The training process characterized "the propagation of noise through the system" to identify and remove it, suggesting a ground truth related to accurate noise modeling and reduction.

Ask a specific question about this device

The Deep Learning Image Reconstruction for Gemstone Spectral Imaging option is a deep learning based CT reconstruction method intended to produce cross-sectional images by computer reconstruction of dual energy X-ray transmission data acquired with Gemstone Spectral Imaging, for all ages. Deep Learning Image Reconstruction for Gemstone Spectral Imaging can be used for whole body, vascular, and contrast enhanced head CT applications.

Deep Learning Image Reconstruction for Gemstone Spectral Imaging (DLIR-GSI) is the next step in CT reconstruction advancement. Like its predicate device (DLIR), DLIR-GSI is an image reconstruction method that uses a dedicated Convolution Neural Network (CNN) that has been designed and trained specifically to reconstruct CT GSI Images to give an image appearance similar to traditional FBP images while maintaining or improving the performance of ASiR-V. The DLIR-GSI can generate monochromatic images (MC), material decomposition images (MD), and virtual unenhanced images (VUE). Multiple MD images such as lodine, Water, Calcium, Hydroxyapatite (HAP), Fat, Uric Acid can be prescribed by the user and generated by the subject device. DLIR-GSI demonstrates same or better Imaging performance as compared to ASiR-V in the following areas: low contrast detectability (LCD), image noise, contrast to noise ratio (CNR), high contrast spatial resolution, CT number accuracy, MD quantification accuracy and metal artifact reduction. Reconstruction times with DLIR-GSI support a normal throughput for routine CT.

The device is marketed as Deep Learning Image Reconstruction for Gemstone Spectral Imaging and the images produced are branded as "TrueFidelity™ CT Images".

Deep Learning Image Reconstruction for Gemstone Spectral Imaging is compatible with dual energy scan modes using the standard kernel and was trained specifically on the Revolution CT family of systems (K163213, K133705, K19177). 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 DLR-GSI: Low, Medium, or High. The strength selection will vary with individual users' preference for the specific clinical need.

As compared to the predicate device, the intended use of Deep Learning Image Reconstruction for Gemstone Spectral Imaging does not change (head and whole-body CT image reconstruction). Both algorithms are designed to produce cross-sectional images of the head and body by computer reconstruction of X-ray transmission data, for all ages.

Acceptance Criteria and Study Details for Deep Learning Image Reconstruction for Gemstone Spectral Imaging (DLIR-GSI)

1. Table of Acceptance Criteria and Reported Device Performance:

The study concluding that the device meets the acceptance criteria is based on:

• Bench Testing: Performed on the identical raw datasets obtained on GE's Revolution CT family of systems, applying both DLIR-GSI and ASiR-V reconstructions for comparison.
• Clinical Reader Study: A retrospective study involving radiologists evaluating images reconstructed with both ASiR-V and DLIR-GSI.

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

• Test Set Sample Size:
  • Main Reader Study: 40 retrospectively collected cases.
  • Additional Clinical Evaluation: 7 additional retrospectively collected clinically challenging cases.
• Data Provenance: Retrospectively collected clinical cases. The country of origin is not specified in the provided text.

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

• Main Reader Study: 5 board-certified radiologists.
  • Qualifications: Expertise in the specialty areas that align with the anatomical region of each case. (e.g., three readers for body/extremity, three for contrast-enhanced head/neck, one qualified for both). Specific years of experience are not mentioned.
• Additional Clinical Evaluation: 1 board-certified radiologist.
  • Qualifications: Expertise in the specialty area that aligns with all cases containing small, low-contrast objects. Specific years of experience are not mentioned.

4. Adjudication Method for the Test Set:

• Main Reader Study: Each image was read by 3 different radiologists. The radiologists provided an assessment of image quality using a 5-point Likert scale.
  • Adjudication Method: Implicitly, a consensus or agreement among the 3 readers would have been used for the assessment of diagnostic quality and noise texture preference. The document states, "The result of this reader study confirmed that the DLIR-GSI (the subject device) produce diagnostic quality images and have preferred noise texture than the reference device ASiR-V," suggesting that the collective findings of the readers led to this confirmation. Explicit details of a 2+1 or 3+1 adjudication are not provided.
• Additional Clinical Evaluation: A single board-certified radiologist evaluated the 7 challenging cases. No adjudication method was applicable as there was only one reader.

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

• Yes, an MRMC study was performed (the "Clinical Testing" described).
• Effect Size (Improvement with AI vs. without AI assistance): The document states that readers confirmed DLIR-GSI produced "diagnostic quality images" and had "preferred noise texture" compared to ASiR-V (considered without AI assistance in this context, or a lesser AI-powered reconstruction). The additional evaluation confirmed "adequately visualized for diagnostic use" in challenging cases. However, specific quantitative effect sizes (e.g., a percentage improvement in diagnostic accuracy, a specific change in AUC, or a numerical metric of improvement in reader performance) are not provided in the given text.

6. Standalone (Algorithm Only) Performance Study:

• Yes, a standalone performance study was done through engineering bench testing. This testing compared DLIR-GSI (algorithm only) against ASiR-V (reference/control algorithm) using identical raw datasets. Metrics like LCD, image noise, CNR, spatial resolution, CT number accuracy, material decomposition accuracy, iodine detection, and metal artifact reduction were evaluated directly from the reconstructed images without human interpretation.

7. Type of Ground Truth Used:

• Bench Testing: The ground truth for metrics like LCD, image noise, spatial resolution, CT number accuracy, etc., would have been based on physical phantom measurements and known parameters of the phantoms used in the engineering tests.
• Clinical Reader Study: The ground truth for image quality and diagnostic usability was established by expert consensus/interpretation from the board-certified radiologists. The text doesn't mention pathology or outcomes data as the primary ground truth for the reader study, but rather the radiologists' assessment of diagnostic quality and visualization.

8. Sample Size for the Training Set:

• The document states that the neural network was "trained specifically to reconstruct CT GSI Images" using "single energy acquired images on the CT Scanner" (for the predicate) and "dual energy acquired images on the CT Scanner" (for the proposed device). It also mentions "information obtained from extensive phantom and clinical data" was used for noise characteristics.
• However, the specific sample size (number of images or cases) for the training set is NOT provided in the text.

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

• The text implies that the neural network was trained to produce an "image appearance similar to traditional FBP images while maintaining or improving the performance of ASiR-V." This suggests that the "ground truth" for training was implicitly the characteristics of high-quality CT images, likely leveraging existing FBP and ASiR-V reconstructed images from "extensive phantom and clinical data."
• For noise modeling, ground truth was based on "characterization of the photon statistics as it propagates through the preprocessing and calibration imaging chain" and using a trained neural network that "models the scanned object using information obtained from extensive phantom and clinical data."
• Specific details on how the ground truth was rigorously established for the training data (e.g., expert annotations, pathology correlation, quantitative metrics derived from known phantoms) are NOT explicitly described. The training appears to be focused on matching or improving upon established reconstruction methods using a large dataset.

Ask a specific question about this device