(100 days)
SwiftMR is a stand-alone software solution intended to be used for acceptance, enhancement and transfer of brain. spine, knee, ankle, shoulder and hip MR images in DICOM format. It can be used for noise reduction and increasing image sharpness for non-contrast enhanced MR images.
SwiftMR is not intended for use on mobile devices.
SwiftMR, is software used as a Medical Device (SaMD) consisting of a software algorithm that enhances images taken by MRI scanners. The device only processes DICOM images for the end User and is intended to be used by radiology technologists in an imaging center, clinic, or hospital.
The device's inputs are standard of care MRI images in DICOM format. The deep learning algorithm produces enhanced images as outputs with reduced noise and increased sharpness in DICOM format. The deep learning algorithm performs the denoising function and sharpening filter performs the sharpening function with the ability of adjusting the sharpness level from level 0 to level 5.
SwiftMR provides an automatic image quality enhancement function for MR images acquired in various environments. SwiftMR can only be used for professional purposes and is not intended for use on mobile devices.
SwiftMR 's automation procedure is as follows:
- . Receive MR images that are in DICOM format
- . Image quality enhancement using Deep Learning model and sharpening filter
- Transfer enhanced MR image as DICOM format to PACS .
There are four deep learning algorithms: three are for the general pulse sequences and the other one is for the TOF pulse sequences. The four deep learning algorithms share the same network architecture, input and label data generation method, training procedures. The only difference between the four deep learning algorithms is the input / label dataset used for training.
After integration with the facilities PACS, SwiftMR performs image processing in the background automatically. At the same time, SwiftMR allows logged-in users to use its functions and view the processing status. When loqged in as the System Admin, the function is available to the control automation procedure and system change settings. On the User side, the User can retrieve the results of image processing in the form of a worklist by login to the user account.
The software provides three main functions, which are image processing, quality check and progress monitoring.
The software is intended to run automatically in the background so that it does not interrupt the workflow of users. When the user executes MR scans as he/she usually does, the newly acquired images are automatically uploaded to the server and registered in the database (DB) for image processing. Once image processing is complete, the images are sent to PACS.
If the user wishes to monitor this automated workflow to check on the status of image processing, he/she can check the main page of the client application or toast messages will appear on the bottom right corner upon completion of each processing. After using the software, they should log out for security reasons.
A settings menu is provided in the form of a user interface to enable system admin to modify software settings as required by the institution or respective user.
Here's a breakdown of the acceptance criteria and the study proving the device meets them, based on the provided FDA 510(k) submission:
1. Table of Acceptance Criteria and Reported Device Performance
| Performance Metric | Acceptance Criteria | Reported Device Performance |
|---|---|---|
| Noise Reduction (SNR) | The average signal-to-noise ratio (SNR) of the SwiftMR-processed image series is increased by 40% or more compared to the value of the original image series. | "This test passed." (Explicitly stated that the acceptance criteria for noise reduction were met.) |
| Sharpness Increase (FWHM) | The Full Width at Half Maximum (FWHM) of a selected region of interest (ROI) is decreased by: |
- 0.43% or more for level 1
- 1.7% or more for level 2
- 2.3% or more for level 3
- 3.6% or more for level 4
- 4.5% or more for level 5
for at least 90% of the dataset. | "This test passed." (Explicitly stated that the acceptance criteria for sharpness increase were met.) |
2. Sample Size Used for the Test Set and Data Provenance
- Sample Size: The document does not explicitly state the exact numerical sample size (number of images or patients) for the validation test set. It mentions the dataset includes various manufacturers (SIEMENS, GE, PHILIPS), field strengths (1.5T/3.0T), anatomical regions (Brain, Spine, Knee, Ankle, Shoulder, Hip), protocols (T1, T2, T2 FS, GRE, FLAIR, PD, PD FS, TOF, SWI, MPRAGE), and demographics (age 22-90, gender Male 48.7%/Female 51.3%). It also states that the validation dataset included data from sources not included in the training dataset to demonstrate performance is not hindered by site variability.
- Data Provenance:
- Country of Origin: Not specified in the provided text.
- Retrospective or Prospective: The study used "retrospective clinical images" for performance testing.
3. Number of Experts Used to Establish the Ground Truth for the Test Set and Qualifications of Those Experts
The document does not specify the number of experts or their qualifications for establishing ground truth for the test set. The acceptance criteria for noise reduction (SNR) and sharpness (FWHM) are quantitative, suggesting these metrics were likely derived through automated or semi-automated image analysis rather than direct expert labeling for "ground truth" in the sense of pathological diagnosis. While experts may have been involved in defining the methodologies for SNR and FWHM calculation or in qualitative assessment, this is not detailed.
4. Adjudication Method for the Test Set
Not applicable/Not described. The assessment relied on quantitative measures (SNR, FWHM) against predefined thresholds, not on expert consensus or adjudication of subjective interpretations.
5. If a Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study was Done
No. The document describes a "Performance test" using retrospective clinical images focused on quantitative metrics (SNR, FWHM) to demonstrate improvements in image quality (noise reduction and sharpness increase). It does not describe an MRMC study comparing human reader performance with or without AI assistance.
6. If a Standalone (i.e. algorithm only without human-in-the-loop performance) was Done
Yes, the performance tests described are standalone evaluations of the algorithm's output. The criteria (SNR increase, FWHM decrease) directly measure the image processing capabilities of SwiftMR without human intervention in the loop for assessment. The device is described as "SwiftMR, is software used as a Medical Device (SaMD) consisting of a software algorithm that enhances images taken by MRI scanners." It operates automatically in the background.
7. The Type of Ground Truth Used
The "ground truth" for the performance evaluation appears to be quantitative, intrinsic image quality metrics:
- Signal-to-Noise Ratio (SNR): A measure of image quality directly related to noise levels.
- Full Width at Half Maximum (FWHM): A measure of image sharpness/resolution.
These metrics are calculated from the images themselves, both original and processed, to quantify the device's effect, rather than based on a diagnostic "ground truth" like pathology outcomes or expert consensus on diagnosis.
8. The Sample Size for the Training Set
The document states, "The only difference between the four deep learning algorithms is the input / label dataset used for training." However, the specific sample size (number of images or patients) for the training set is not provided.
9. How the Ground Truth for the Training Set was Established
The document states that the deep learning algorithms were trained using "input and label data generation method." The "labels" in this context for image enhancement typically refer to reference or target images. It specifies that "Neural network-based filters that simultaneously perform noise reduction and sharpness increase functions are obtained. The parameters of the filters were obtained through an image guided optimization process." For deep learning-based image enhancement, the "ground truth" or "labels" for training often involve:
- Synthetically generated data: Creating pairs of noisy/blurry images and their corresponding "clean/sharp" counterparts.
- Paired clinical data: Acquiring both lower quality (e.g., fast scan) and higher quality (e.g., long scan, high-resolution) images of the same anatomy, with the higher quality images serving as the "ground truth" or target for the lower quality input.
The document does not provide explicit details on the specific method of "label data generation" for training, beyond mentioning an "image guided optimization process."
§ 892.2050 Medical image management and processing system.
(a)
Identification. A medical image management and processing system is a device that provides one or more capabilities relating to the review and digital processing of medical images for the purposes of interpretation by a trained practitioner of disease detection, diagnosis, or patient management. The software components may provide advanced or complex image processing functions for image manipulation, enhancement, or quantification that are intended for use in the interpretation and analysis of medical images. Advanced image manipulation functions may include image segmentation, multimodality image registration, or 3D visualization. Complex quantitative functions may include semi-automated measurements or time-series measurements.(b)
Classification. Class II (special controls; voluntary standards—Digital Imaging and Communications in Medicine (DICOM) Std., Joint Photographic Experts Group (JPEG) Std., Society of Motion Picture and Television Engineers (SMPTE) Test Pattern).