Search Results

HYPER DLR is an image processing function intended to be used by radiologists and nuclear medicine physicians to reduce noise of the fluorodeoxyglucose (FDG) PET images.

HYPER DLR is a software-only device. HYPER DLR is intended to be implemented on previously cleared PET/CT devices uMI 550 (K182237) and uMI 780 (K172143). HYPER DLR serves as an alternative to the existing image smoothing options that are available on the predicate devices. HYPER DLR is an image post-processing technique which uses a pre-trained neural network to predict low noise PET image from high noise PET image. After training, the network could extract the noise component from the image, thus reducing the image noise.

Here's a breakdown of the acceptance criteria and the study information for the HYPER DLR device, based on the provided FDA 510(k) summary:

Acceptance Criteria and Reported Device Performance

The document describes the device performance in qualitative terms and through common quantitative metrics, but does not provide a specific table of numerical acceptance criteria with corresponding performance values for clinical image evaluation. The acceptance for clinical evaluation is described as:

For non-clinical (bench) testing, the document states: "Bench test showed overall image quality improvement based on the commonly used quantitative metrics. HYPER DLR can significantly improve SNR and CNR while preserving image consistency." The specific acceptance values for "significant improvement" are not detailed in this summary. The quantitative metrics evaluated include:

• Peak signal to noise ratio
• Structural similarity index
• Pearson correlation coefficient
• Signal to noise ratio (SNR)
• Contrast to noise ratio (CNR)
• Normalized root mean square error
• Bland-Altman plot of body & brain VOI SUVmean values

Study Details

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

   • Test Set Sample Size: Not explicitly stated as a number of cases or images. The document mentions "clinical image evaluation were performed for typical clinical scan times of uMI 550 and uMI 780 systems." This implies a set of clinical images, but the exact count is not provided.
   • Data Provenance: Not explicitly stated. The manufacturer is Shanghai United Imaging Healthcare Co., Ltd. in China, so it's plausible the data originates from studies conducted there or affiliated sites. The document does not specify if the data was retrospective or prospective.

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

   • Number of Experts: Not explicitly stated as a specific number. The document mentions "Each image was read by board-certified nuclear medicine physicians." It implies multiple physicians but doesn't specify how many.
   • Qualifications: Board-certified nuclear medicine physicians.

3. Adjudication method for the test set:

   • The document does not explicitly describe an adjudication method for disagreements among the physicians. It states that physicians "provided an assessment," implying individual assessments that were then analyzed.

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:

   • A multi-reader, multi-case study was done for clinical image evaluation, comparing HYPER DLR images with Gaussian filtered images. However, this study appears to be a standalone reader study comparing two different image processing methods, not an AI-assisted vs. non-AI-assisted human reader study. Therefore, no effect size for human reader improvement with AI assistance is reported because the study design was different. The physicians were evaluating images generated by the AI algorithm (HYPER DLR) versus images generated by conventional post-smoothing (Gaussian filtering).

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

   • Yes, the "Bench test" section describes standalone algorithm performance evaluation using quantitative metrics like SNR, CNR, etc. This solely assesses the algorithm's output characteristics without human interpretation.

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

   • For the clinical image evaluation, the ground truth was based on the "assessment of both image noise and overall image quality" by board-certified nuclear medicine physicians. This could be interpreted as a form of expert consensus or expert opinion on image characteristics rather than a definitive "true positive/negative" ground truth for disease detection, as the device's purpose is noise reduction and image quality improvement.
   • For the bench testing, the ground truth for quantitative metrics would typically be derived from the inherent properties of the phantoms/datasets used for those measurements (e.g., known signal levels, known noise levels).

7. The sample size for the training set:

   • Not explicitly stated. The document mentions that HYPER DLR "uses a pre-trained neural network," but the size of the dataset used for this training is not disclosed in this summary.

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

   • Not explicitly stated in this summary. For deep learning noise reduction, the training process often involves pairs of 'noisy' and 'clean' images, where the 'clean' image serves as the ground truth for the noise reduction task. These 'clean' images might be generated from higher dose acquisitions or simulated to represent noise-free data.

Ask a specific question about this device