Clinicians and scientists interested in contact-free pulse triggering methods.Recent studies have shown that video recordings of the human skin, obtained with MRI-compatible cameras, contain a photoplethysmographic signal component that resembles the one measured by pulse oximetry [1,2]. This contact-free measurement could overcome the limitations of contact-based electrocardiography and pulse oximetry: magnetohydrodynamic interferences at ultra-high field strengths, limited suitable body areas, and susceptibility to motion artefacts. Previous works were limited to the rather low temporal resolution of currently available MRI-compatible cameras (25-30 frames-per-second corresponding to a delay of 33-40ms). However, for cardiac MRI triggering, a delay after the electrocardiography R-wave of less than 20ms is recommended [3]. Additionally, only the distinct systolic peak of the photoplethysmography signal was considered in previous works. The purpose of this work was (1) to evaluate the triggering performance of a high-speed camera prototype that fulfills the temporal resolution required for cardiac MRI and (2) to investigate whether more subtle components of the photoplethysmography waveform can be extracted.

Experiments were conducted on two healthy subjects (male, age: 27-35yrs, weight: 65-70kg, height: 172-178cm) in an ultra-high-field scanner room (Magnetom 7T, Siemens GmbH, Erlangen, Germany). As the high-speed camera was a prototype and not approved for high field strengths, subjects lay on the patient table in the home position at approximately 0.3T. A spotlight (MultiLED LT, GSvitec, Gelnhausen, Germany; 7700 Lumen) was used to provide illumination from outside the scanner room. To simplify the manual definition of a region-of-interest (ROI) during video processing, a cardboard stencil (8x3cm cut-out) was attached to the volunteer’s forehead. The described set-up can be used in the MR bore without modifications.

For each volunteer, signals were recorded for 9s duration. The camera used in our previous work (C1: 12M-i, MRC Systems, Heidelberg, Germany; 1/3'' CMOS sensor) as well as a high-speed camera prototype (C2: MRC HiSpeed camera, same vendor; 1/2'' CMOS sensor) were installed above the volunteer’s head using a custom-built stand. Time-stamped frames were acquired using synchronized C++ algorithms (C1: OpenCV library, 30Hz, 720x576 pixel; C2: GigE Vision SDK, 250Hz, 640x480 pixel) executed on an off-the-shelf laptop. Pulse oximetry data was obtained by the MR scanner’s physiological monitoring unit and was not synchronized.

Processing of both video signals was performed offline using Matlab (MathWorks, Natick, MA, USA). A ROI covering 200x200 pixels was manually defined using the stencil as a reference. A simple filter algorithm was used for extracting physiological information from mean pixel values (Fig.1, black) inside the ROI of sequential frames, followed by peak detection: At each mean pixel value, a linear polynomial curve fit was applied using all preceding values, weighted by the left half of a Gaussian window (full width at half maximum: 300ms). The obtained slope signal (Fig.1, blue) was then filtered with a running maximum window of width 300 ms. A trigger was stored if the running maximum signal was constant for three samples. The minimum time until a next trigger could be stored was set to 500ms.

Fig.1 shows the pulse oximeter photoplethysmogram, the mean pixel intensity signals, and the filtered signals for camera C1 and C2. The filtered signal shows a significantly higher agreement with the contact-based photoplethysmogram waveform; the diastolic and systolic peaks as well as the dicrotic notch can be identified. Using camera C2, the higher number of available control points makes curve fitting more robust and the filtered signal is similar smooth as in the case of C1. Fig.2 zooms in on one cardiac cycle. Using the proposed trigger / peak detection algorithm, the high-speed camera allows the detection of triggers earlier than the low-speed camera. Analysis of the experimental data showed that this was the case in 94% of all cardiac cycles, with a mean speed increase of 126ms.

Recently, it has been shown in non-MRI-related work that it is possible to recover the whole photoplethysmography waveform from videos using sophisticated offline algorithms [4]. So far, only the systolic peak was considered in MR-related works [1,2]. The results of the current work suggest that it is possible to measure the diastolic peak and dicrotic notch as well using a simple and potentially real-time filter method. Additionally, we evaluated the performance of a high-speed camera for triggering based on this filter technique followed by peak detection and measured a speed increase compared to conventional cameras that could fulfill the requirements for cardiac MRI. However, considering the small sample size and no synchronization to electrocardiography, more experiments are needed and the validation of real-time feasibility remains to be demonstrated.