(119 days)
QFR is indicated for use in clinical settings where validated and reproducible quantified results are needed to support the assessment of coronary vessels in X-ray angiographic images, for use on individual patients with coronary artery disease.
When the quantified results provided by QFR are used in a clinical setting on X-ray images of an individual patient. The results are only intended for use by the responsible clinicians.
QFR is delivered as a standalone software package which is installed and running on a server system in the server room of the cathlab or the hospital. The server offers all functionalities that are required to work with the quantitative measurement in X-ray Angiographic (XA) patient studies supported by the QFR device.
QFR will be used by interventional cardiologists and researchers to obtain quantifications of lesions in coronary vessels. QFR has been developed as a web-based application to run in a web browser in the control room of the cathlab or in a hospital image review room. The import of images and the export of analysis results are via PACS.
The QFR device calculates the QFR value based on an anatomical model which is the result of a 3D reconstruction using the 2D contours obtained from two angiographic projections with angles >=25 degrees apart. These projections are acquired through monoplane or biplane XA systems. The algorithm involves three key steps: (1) Vessel Selection, (2) Contours Detection, and (3) QFR Analysis:
-
Vessel Selection: Angiograms are pre-classified by a deep learning model, identifying main epicardial vessels such as RCA, LAD, and LCx. The user then chooses the segment for analysis, and the software automatically selects end-diastolic image frames. The end-diastolic frame is determined as the angiogram frame with the vessel lumen adequately filled with contrast in both image sequences. This selection is either based on the patient's electrocardiogram when available or performed by the software using a deep learning model. It is essential for the user to verify this selection before proceeding with the analysis. The chosen end-diastolic frame serves as the projection view for the subsequent 3D reconstruction of the vessel.
-
Contours Detection. First, the system runs another deep learning model for coronary vessel segmentation as input to identify anatomical corresponding points on both projections for automatic correction of the system distortions introduced by the isocenter offset and the respiration-induced heart motion. Second, begins the automatic detection of start and end positions of the vessel segment to be reconstructed on the projection views, and extract its contours and centerline. Third, the position of the start and end point must be confirmed by the user.
-
QFR Analysis: The QFR value is computed from the arterial and reference diameter function calculated from the 3D reconstruction based on the contours detected on the cross-sections of the vessel segment, and the patient-specific volumetric flow rate calculated from the automated TIMI frame count. The reference diameter and bifurcations are used to determine the flow distribution at coronary bifurcations and calculate the reference diameter function. The reconstructed 3D model is used to calculate the QFR value.
A report is generated by QFR that shows patient information, image acquisition information (both obtained from the DICOM input), analysis results (vessel sizing and QFR value) and snapshot images showing the vessel boundaries.
The provided FDA 510(k) Clearance Letter for QFR (3.0) outlines the device's acceptance criteria and performance data from a study. Here's a breakdown of the requested information:
1. Table of Acceptance Criteria and Reported Device Performance
The acceptance criteria are not explicitly stated in a single, clear table. Instead, they are defined for specific algorithmic improvements. The reported performance is then compared against these defined criteria.
| Feature Evaluated | Acceptance Criteria | Reported Device Performance | Result (Met/Not Met) |
|---|---|---|---|
| Vessel Classification | 80% for correct vessel classification (since it supports the user, not fully automates) | RCA: 96% correct; LAD: 88% correct; LCx: 78% correct. "On average the 80% acceptance criterion was satisfied." | Met |
| Start and End Point Detection | 80% for correct result, with only 10% allowed to be completely wrong (since it supports the user, not fully automates) | AI/ML model using image data: 77% correct result, 11% small deviation (needed no correction), 8% wrong result (needed correction), 4% gave no result. "In conclusion, 88% satisfies the 80% criterion and 8% satisfies the 10% criterion." (Note: The 88% is derived from 77% correct + 11% small deviation which needed no correction. The 8% wrong result is within the 10% allowed. The 4% "no result" is not explicitly addressed by criteria but the overall conclusion is favorable.) | Met |
| End-Diastolic (ED) Frame Detection | 80% for correct detection of the ED (since it supports the user, not fully automates) | Analytical algorithm (using ECG data): 83% on a representative dataset. AI/ML model (using image data): 81% on a representative dataset. | Met |
| New Flow Velocity Calculation (influencing QFR) | Acceptance criterion "significantly stricter for the resulting QFR measurement than the reproducibility of FFR measurements." | "This ensured that, the automatic flow calculation, was not outperformed by manual indication." (The specific numerical values of the "stricter" criterion and validation results are not provided, only a general statement of meeting the intent.) | Met (implied) |
2. Sample Size Used for the Test Set and the Data Provenance
- Test Set Sample Size: For the "Vessel classification," "Start and end point detection," and "End-Diastolic (ED) frame detection" evaluations, the document mentions being performed "on a representative dataset." However, the exact sample size (number of patients or images) for these test sets is not explicitly stated.
- Data Provenance: The document does not specify the country of origin of the data or whether the data was retrospective or prospective.
3. Number of Experts Used to Establish the Ground Truth for the Test Set and the Qualifications of Those Experts
The document does not explicitly state the number of experts used to establish the ground truth for the test set or their specific qualifications (e.g., "radiologist with 10 years of experience"). The context implies that for functionalities "supporting the user and not to completely automate the functionality," human review and correction are part of the process, suggesting expert involvement, but the formal ground truth establishment process is not detailed.
4. Adjudication Method (e.g., 2+1, 3+1, none) for the Test Set
The document does not specify an adjudication method (such as 2+1 or 3+1) for establishing the ground truth for the test set.
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
The document does not describe a formal Multi Reader Multi Case (MRMC) comparative effectiveness study designed to measure the improvement of human readers with AI assistance versus without AI assistance. The performance evaluations stated are for the algorithm's ability to assist (e.g., correct classification or detection rates) rather than human performance metrics. The statement "For all of these algorithmic improvements the user is able to review and correct the results before the QFR value is calculated" implies that the AI is assistive, but no data on human performance improvement is presented.
6. If a Standalone (i.e. algorithm only without human-in-the-loop performance) Was Done
Yes, standalone performance of the algorithm components (e.g., vessel classification, start/end point detection, ED frame detection by AI/ML model alone) was evaluated and reported against the acceptance criteria. For example, for vessel classification, 96% correct for RCA, 88% for LAD, and 78% for LCx are standalone algorithmic performance numbers before human review and correction. Similarly, the 77% correct for start/end point detection and 81% for ED frame detection are standalone algorithmic performances.
7. The Type of Ground Truth Used (expert consensus, pathology, outcomes data, etc.)
The document implies that the ground truth for the "correct" classifications/detections was based on some form of human reference standard or expert review, as the system is designed to "support the user" and allows for "manual correction." However, the specific method of establishing this ground truth (e.g., expert consensus, comparison to a gold standard, or clinical outcomes) is not explicitly detailed. For the flow velocity calculation influencing QFR, the ground truth is implicitly related to QFR results and their validation against FFR (Fractional Flow Reserve) reproducibility, which is a physiological measurement.
8. The Sample Size for the Training Set
The document does not specify the sample size for the training set used for the deep learning models. It only mentions that the angiograms are "pre-classified by a deep learning model" and that the system "runs another deep learning model for coronary vessel segmentation."
9. How the Ground Truth for the Training Set Was Established
The document does not explicitly describe how the ground truth for the training set was established. It is implied that for deep learning models, labeled data would have been required, but the process of labeling (e.g., by experts, automated methods, or a combination) is not detailed.
§ 892.1600 Angiographic x-ray system.
(a)
Identification. An angiographic x-ray system is a device intended for radiologic visualization of the heart, blood vessels, or lymphatic system during or after injection of a contrast medium. This generic type of device may include signal analysis and display equipment, patient and equipment supports, component parts, and accessories.(b)
Classification. Class II.