(273 days)
DeepFoqus-Accelerate is a stand-alone software solution intended to be used for acceptance, enhancement, and transfer of brain MRI images in DICOM format. It can be used for reconstruction of non-contrast enhanced MRI images acquired with 1.5T or 3T Siemens and GE scanners using Sagittal, Axial, or Coronal T1, T2, or FLAIR sequences.
DeepFoqus-Accelerate is intended to be used on adult scans only and not intended for use on mobile devices.
DeepFoqus-Accelerate is an AI-powered MRI software that accepts up to 4x accelerated scans and reconstructs them to yield acceptable scans when compared to standard unaccelerated scans.
DeepFoqus allows input of HDF5 and DICOM files, acquired with common MRI machines, containing accelerated T1, T2, and FLAIR image sets. Once inputted, DeepFoqus performs a series of reconstruction steps to produce clinical-quality scans from the undersampled data in the phase encoding direction.
The full pipeline involves the following steps:
- Ingestion of MRI scans as k-space and DICOM files.
- Preprocessing steps to ensure that the input is in a standard format and meets acquisition requirements.
- Processing through a pipeline that combines a collection of machine learning and signal processing modules that ensures that an appropriate range of fine-tuned models is applied to reconstruct a final volume.
- Post-processing steps which include adjustments for windowing and image size.
- Conversion of the final volume to DICOM format.
The software provides a workflow for an MRI/Radiology Technologist to:
- Input undersampled data into the software through a user interface
- Start a processing pipeline and observe progress
- Export reconstructed data for use in downstream applications
The provided text describes the FDA 510(k) clearance for DeepFoqus (DeepFoqus-Accelerate), an AI-powered MRI software. While the document outlines various tests and validations performed, it does not explicitly list quantitative acceptance criteria with corresponding reported device performance values for a clinical performance study in a table format, nor does it provide detailed information on sample sizes, expert qualifications, or adjudication methods for the clinical validation beyond general statements.
However, based on the information provided, here's an attempt to answer the questions, highlighting where specific details are missing and assumptions are made based on typical regulatory submissions:
Acceptance Criteria and Device Performance
The document broadly states that "Software Validation" included "Analytical performance validation" and "Clinical performance validation." For clinical performance, it mentions: "Radiologists reviewed a sample of diverse accelerated scans (including pathological and non-pathological brains) to determine whether the accelerated scans were equivalent to the ground truth."
Without specific numerical acceptance criteria and corresponding performance metrics from the provided document, a table cannot be fully constructed. The document only lists the metrics used for analytical performance.
Placeholder Table (Based on typical expectations for such devices, as specific quantitative criteria are NOT provided in the document):
| Performance Metric | Acceptance Criteria (Assumed/Typical) | Reported Device Performance (Not specified in document) |
|---|---|---|
| Analytical Performance | ||
| SSIM (Structural Similarity Index Measure) | Maintain similarity > X (e.g., 0.95) compared to unaccelerated scans | "demonstrate acceptable performance and equivalence with the predicate device" (No specific value given) |
| PSNR (Peak Signal to Noise Ratio) | Maintain PSNR > Y (e.g., 30 dB) compared to unaccelerated scans | "demonstrate acceptable performance and equivalence with the predicate device" (No specific value given) |
| HaarPSI (Haar wavelet-based perceptual similarity index) | Demonstrate perceptual similarity (e.g., 0.9) compared to unaccelerated scans | "demonstrate acceptable performance and equivalence with the predicate device" (No specific value given) |
| Clinical Performance (Qualitative Radiologist Review) | ||
| Equivalence to Ground Truth (Radiologist Assessment) | Accelerated scans are determined to be "equivalent" to ground truth in a majority of cases by expert radiologists. | "Radiologists reviewed...to determine whether the accelerated scans were equivalent to the ground truth." (No specific percentage or rating given) |
| Non-inferiority to Predicate | Performance outcomes (e.g., image quality, diagnostic utility) are non-inferior to the predicate device. | "performs comparably to the predicate device" |
Study Details
2. Sample Size and Data Provenance
- Test Set (Clinical Performance Validation): The document states "Radiologists reviewed a sample of diverse accelerated scans (including pathological and non-pathological brains)".
- Sample Size: Not specified.
- Data Provenance: Not explicitly stated (e.g., country of origin). It mentions "diverse" datasets, which could imply multi-center or varied origins, but this is not confirmed.
- Retrospective/Prospective: Not specified. Typically, for such validation, retrospective datasets are used.
3. Number of Experts and Qualifications for Ground Truth
- Number of Experts: Not specified. It only states "Radiologists reviewed a sample".
- Qualifications of Experts: Not specified beyond "Radiologists". Typical expectations are board-certified radiologists with experience in neuro-imaging.
4. Adjudication Method for the Test Set
- Adjudication Method: Not specified. It only states "Radiologists reviewed". Common methods include consensus reading, majority vote (e.g., 2+1, 3+1), or independent readings.
5. Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study
- Was an MRMC study done? The document describes a "Clinical performance validation" where "Radiologists reviewed" scans to determine their equivalence to ground truth. It also states that the device "performs comparably to the predicate device."
- It doesn't explicitly state if this was formalized as a multi-reader, multi-case comparative effectiveness study with and without AI assistance to measure human reader improvement. It focuses on the device's output "equivalence" to ground truth and its comparability to the predicate.
- Therefore, the direct effect size of how much human readers improve with AI vs. without AI assistance is not provided or explicitly described as a primary endpoint. The device acts as a reconstruction tool for accelerated scans, aiming to produce clinical-quality images from undersampled data, rather than directly assisting in diagnosis of an already acquired scan.
6. Standalone (Algorithm Only) Performance
- Was a standalone study done? Yes, implicitly. The "Analytical performance validation" using metrics like SSIM, PSNR, and HaarPSI directly assesses the algorithm's output against a reference (standard unaccelerated scans). This is algorithm-only performance, as it doesn't involve human interpretation to generate these metrics.
- The clinical performance validation also seems to assess the output of the algorithm for "equivalence to ground truth" via radiologist review, which can be considered a standalone assessment of the reconstructed images.
7. Type of Ground Truth Used
- Ground Truth Type:
- For analytical performance, the ground truth was "standard unaccelerated scans". This can be considered a reference standard image.
- For clinical performance, "ground truth" was established by comparing the reconstructed accelerated scans to "standard unaccelerated scans" as reviewed by radiologists. This implies the non-accelerated, conventionally acquired MRI served as the clinical reference for quality and diagnostic interpretability. The document mentions "pathological and non-pathological brains", suggesting the ground truth also encompassed known clinical findings.
8. Sample Size for the Training Set
- Training Set Sample Size: Not specified.
- However, the document states: "Training has been conducted on a range of datasets which include diverse magnet strengths, sequences, manufacturers, and image orientations." This indicates a broad and varied training corpus.
9. How Ground Truth for Training Set was Established
- Ground Truth for Training Set: The document states the "overall pipeline has been trained to minimize the distance between reconstructed final images and standard accelerated scans." This implies that pairs of accelerated (undersampled) and "standard unaccelerated" MRI scans were used for training.
- The "standard unaccelerated scans" served as the reference or ground truth for the Deep Learning models to learn to reconstruct the accelerated data. This is a common supervised learning approach in image reconstruction, where the model learns to map an undersampled input to a fully sampled, high-quality output.
§ 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).