Quantitative quality-control metrics for in vivo oximetry in small vessels by visible light optical coherence tomography angiography

Abstract

Biological functions rely on local microvasculature to deliver oxygen and nutrients and carry away metabolic waste. Alterations to local oxygenation levels are manifested in diseases including cancer, diabetes mellitus, etc. The ability to quantify oxygen saturation (sO₂) within microvasculature in vivo to assess local tissue oxygenation and metabolic function is highly sought after. Visible light optical coherence tomography (vis-OCT) angiography has shown promise in reaching this goal. However, achieving reliable measurements in small vessels can be challenging due to the reduced contrast and requires data averaging to improve the spectral data quality. Therefore, a method for quality-control of the vis-OCT data from small vessels becomes essential to reject unreliable readings. In this work, we present a quantitative metrics to evaluate the spectral data for a reliable measurement of sO₂, including angiography signal to noise ratio (SNR), spectral anomaly detection and discard, and theory-experiment correlation analysis. The thresholds for each quantity can be flexibly adjusted according to different applications and system performance. We used these metrics to measure sO₂ of C57BL/6J mouse lower extremity microvasculature and validated it by introducing hyperoxia for expected sO₂ changes. After validation, we applied this protocol on C57BL/6J mouse ear microvasculature to conduct in vivo small blood vessel OCT oximetry. This work seeks to standardize the data processing method for in vivo oximetry in small vessels by vis-OCT.

Fig. 1

Schematic and scanning protocols of the visible light spectroscopic optical coherence tomography (vis-OCT) imaging system. (a) Illustration of a blood vessel embedded in tissue. (b) Optical setup of vis-OCT. SL: supercontinuum source; DM: dichroic mirror; PBS: polarization beam splitter; DSM: D-shaped mirror; P: prism; B: block; M: mirror; BT: beam trap; L: lens (f = 10 mm); SPEC: spectrometer; OFC: wide band optical fiber coupler; PC: polarization controller; VNDF: variable neutral density filter; DC: dispersion control; GV: galvanometer mirror; OL: objective lens (Effective working length: 39 mm). (c) Illustration of scanning protocol 1. (d) Illustration of scanning protocol 2.

Fig. 2

Flow chart of the three-step statistical data-cleaning process of vis-OCT raw spectra. SNR: signal to noise ratio;$x_i$: the$i^{th}$spectrum of all repetitive spectra of a blood vessel;$\mu$: the mean spectra of all repetitive B-scans of a blood vessel;$d(x_i,\mu )$: the Mahalanobis distance (MD) between$x_i$and$\mu$;$\theta$: the intersection angle between simulated and vis-OCT measured spectra of a blood vessel; sO₂: oxygen saturation.

Fig. 3

Data quality control of vis-OCT spectra from six different sized blood vessels. (a) The en face angiography of wild-type C57BL/6J mouse lower extremity microvasculature acquired by scanning protocol 1. The red vertical line marks the location of a B-scan to be repetitively scanned. (b) The en face angiography of the marked B-scan in (a) acquired by scanning protocol 2. The color bar encodes depth locations of vessel central axis within 100$\mu m$ from sample top surface. Scale bar: 100$\mu m$. Correspondence to the same blood vessels shown in (a) and (b) is indicated by red arrows in between. Blood vessels are numbered as Vessel 1, Vessel 2, Vessel 3, Vessel 4, Vessel 5, and Vessel 6, respectively. (c) The CV of $L({\mu }_n)$, the mean MD between the validation samples and randomly selected training samples, for six vessels, respectively. V1 - V6: Vessel 1 to Vessel 6. The red dashed line indicates CV = 5%. The black box zooms in on the region where CVs are around 5%. (d) – (i) The vis-OCT measured spectra processed by the data-cleaning and their corresponding simulated oxygenated, deoxygenated, and sO₂ fitted spectra for V1 to V6, respectively. Simulated spectra are normalized by the maximum of oxygenated spectra within 555 – 572 nm, indicated by black vertical dashed lines; vis-OCT measured spectra are scaled to have the same mean as sO₂ fitted spectra. Sim. O/D: simulated oxygenated/deoxygenated spectra; Exp.: Experimental measurements; sO₂ fit: sO₂ fitted spectra; Org./Cld CV: the coefficients of variation (CV) of the averaging OCT spectra before (original) and after (cleaned) the data-cleaning process. CV = S.D./Mean.

Fig. 4

Validation of the data quality-control protocol for in vivo vis-OCT Oximetry by supplying normal air and 100% oxygen to a wild-type C57BL/6J mouse. (a) The en face angiography of the mouse lower extremity microvasculature at normal air condition. (b) The en face angiography of the same imaging site in (a) ventilated by 100% pure oxygen. In (a) and (b), A: arteriole; V: venule; 3-9: vessel 3 to vessel 9. The color bar encodes depth of vessel central axis within 100 $\mu m$ from sample top surface. Scale bar: 100$\mu m$. (c) The sizes of blood vessels in (a) and (b) at normal and 100% oxygen conditions in descending order. (d) The quantified sO₂ of blood vessels in (a) and (b) at normal and 100% oxygen conditions, respectively. In (c) and (d), A: arteriole; V: venule; V3 – V9: vessel 3 to vessel 9.

Fig. 5

The en face sO₂ map of C57BL/6J mouse ear microvasculature. The blue arrows indicate locations of the arteriole and the venule. A: arteriole; V: venule. Scale bar: 100$\mu m$.

Fig. 6

Simulation results of OCT spectra in log scale for oxygenated and deoxygenated blood vessels with diameters of 10$\mu m$, 30$\mu m$, 70$\mu m$, and 150$\mu m$, respectively. O: oxygenated state; D: deoxygenated state.

Fig. 7

Scatter plots of Vessel 1 in Fig. 3 at iterations 1, 4, 7, and 11 during the anomaly detection and outlier removal process. Red lines indicate the thresholds for the Mahalanobis distance (MD) between each repetitive spectrum of the blood vessel ($x_i$) and their mean ($\mu$), represented by$\hat{d}=\mu +1.645\sigma$, where$\sigma$ is variance. The values of $\hat{d}$at iterations 1, 4, 7, and 11 are 4.9483, 4.4383, 4.3801, and 4.3672, respectively. Blue dots above the red lines at each iteration were discarded as outliers. The number of data points with MD lower than the thresholds at iterations 1, 4, 7, and 11 are 463, 373, 340, and 332, respectively.

Fig. 8

The number of repetitive B-scans detected as normal at each iteration of the anomaly detection and outlier removal process for (a): Vessel 1, (b): Vessel 2, (c): Vessel 3, (d): Vessel 4, (e): Vessel 5, and (f): Vessel 6 shown in Fig. 3. The total iteration numbers of vessel 1 to vessel 6 are 12, 15, 19, 9, 28, and 35, respectively.

Fig. 9

The sO₂ histograms of blood vessels in Figs. 4(a) and (b) without and with the data-quality process at normal (a) and 100% oxygen (b) conditions.

Fig. 10

The en face sO₂ map of C57BL/6J mouse ear microvasculature without data quality control process. Scale bar: 100$\mu m$.

Fig. 11

The spectroscopic angiography image SNR of the six vessels marked in Figs. 3(a) and 3(b). V1-V6: Vessel 1 to Vessel 6.

Fig. 12

The intersection angles between simulated OCT spectra with and without different zero-mean Gaussian white noise for three different spectral bands. The standard deviations (${\sigma }_G$) of the Gaussian white noise were approximately 10%, 5%, and 2%, respectively, of the mean of the simulated spectrum ($\mu$) with sO₂ as$\alpha$. The oxygen saturations of the simulated OCT spectra are (a): 60%, (b): 70%, (c): 80% and (d): 90%. The three spectral bands are 546 nm - 584 nm, 549 nm - 578 nm, and 552 nm - 572 nm, indicated by red, green, and blue, respectively. All curves are the mean of 100 iterations of the simulations.

Fig. 13

The intersection angles between simulated spectra of blood vessels with sO₂ as $\alpha$ and $\alpha \pm 5\%$ added by a zero-mean Gaussian white noise. The standard deviations (${\sigma }_G$) of the Gaussian white noise were approximately 10%, 5%, and 2% of the mean of the simulated spectrum ($\mu$) with sO₂ as$\alpha$. (a): $\alpha$ = 60%, (b): $\alpha$ = 75%, and (c): $\alpha$ = 90%. All curves are the mean of 100 iterations of the simulations.