Quantitative real-time pulse oximetry with ultrafast frequency-domain diffuse optics and deep neural network processing

Abstract

Pulse oximetry is a ubiquitous optical technology, widely used for diagnosis and treatment guidance. Current pulse oximeters provide indications of arterial oxygen saturation. We present here a new quantitative methodology that extends the capability of pulse oximetry and provides real-time molar concentrations of oxy-and deoxy-hemoglobin at rates of up to 27 Hz by using advanced digital hardware, real-time firmware processing, and ultra-fast optical property calculations with a deep neural network (DNN). The technique utilizes a high-speed frequency domain spectroscopy system with five frequency-multiplexed wavelengths. High-speed demultiplexing and data reduction were performed in firmware. The DNN inversion algorithm was benchmarked as five orders of magnitude faster than conventional iterative methods for optical property extractions. The DNN provided unbiased optical property extractions, with an average error of 0 ± 5.6% in absorption and 0 ± 1.4% in reduced scattering. Together, these improvements enabled the measurement, calculation, and real-time continuous display of hemoglobin concentrations. A proof-of-concept cuff occlusion measurement was performed to demonstrate the ability of the device to track oxyand deoxy-hemoglobin, and measure quantitative photoplethysmographic changes during the cardiac cycle. This technique substantially extends the capability of pulse oximetry and provides unprecedented real-time non-invasive functional information with broad applicability for cardiopulmonary applications.

Frequency-Domain Techniques for Cerebral and Functional Near-Infrared Spectroscopy

“…Recent developments in FD-NIRS instrumentation include implementations based on a compact frequency-sweeping circuit board (No et al, 2008), digital heterodyning (Arnesano et al, 2012), direct digital sampling (Roblyer et al, 2013;Zimmermann et al, 2016), frequency division multiplexing (Torjesen et al, 2017;Zhao et al, 2018), CMOS integrated circuitry (Sthalekar and Joyner Koomson, 2013;Yun and Joyner Koomson, 2013), and vertical cavity surface emitting lasers (VCSEL) (Sultan et al, 2013;Kitsmiller et al, 2018).…”

Recent Developments in Instrumentation of Functional Near-Infrared Spectroscopy Systems

“…In the last five years, there has been renewed interest in improving the fNIRS technology itself leading to new advantages associated with FD and potentially new applications. Roblyer and his team have recently developed an ultrafast frequency-domain diffuse optics system with a deep neural network (DNN) processing method to measure the optical properties [86]. The DNN is used to replace the time-consuming Levenberg-Marquardt iterative algorithm which was adopted to fit the calibrated amplitude and phase measurements to an analytical forward model.…”

“…In contrast to the iterative algorithm, the DNN is 3-5 orders of magnitude faster to estimate the optical properties of measured tissue. Therefore, the developed system combined with DNN enables a robust tissue oxygenation monitoring system that can be able to acquire, process, and display absolute concentrations of hemoglobin at an adequate rate to catch the cardiac cycle at the higher speed [86].…”

Modulation frequency selection and efficient look-up table inversion for frequency domain diffuse optical spectroscopy

“…In between these two extremes, other groups have used two, 12 or several dozen modulation frequencies to measure μ a and μ 0 s . 13 Previous work has investigated the theoretical limits of spatial resolution and sensitivity of FD-DOS measurements for a given single modulation frequency, 14 and there have been investigations to determine an optimal modulation frequency for tomographic reconstruction assuming a shot-noise limited system. [15][16][17] These prior works agree that, if a shot-noise limited system is available, an optimal modulation frequency exists between 300 and 800 MHz depending on the OPs of the sample and the source-detector separation employed.…”

“…The default options have been used extensively in the previous work. 13,[23][24][25] We also implemented an additional check that excluded reconstructed OP pairs that did not meet the μ 0 s > 10μ a criteria. This check was performed because the LMA algorithm would occasionally converge in the region where μ a and μ 0 s were on the same order of magnitude.…”