CC BY
5DOI 10.24412/cl-37136-2023-1-16-18
MACHINE LEARNING-BASED RECONSTRUCTION OF BLOOD OXYGEN SATURATION:
PILOT STUDY
ALEKSANDR KHILOV1, DARIA KURAKINA1, MIKHAIL KIRILLIN1 AND VALERIYA
PEREKATOVA1
1 Federal Research Center A.V. Gaponov-Grekhov Institute ofApplied Physics of the Russian Academy of
Sciences, Russia
ABSTRACT
The problem of the determination of blood oxygen saturation (sO2) is important for a number of biomedical applications [1]. Methods of diffuse optical spectroscopy (DOS) are traditionally used for sO2 measurements. DOS employs the difference in spectra of oxy- and deoxyhemoglobin at two probing wavelengths. However, the resolution of this method is not vessel level, but for particular applications, for example, tumor response to treatment monitoring [2], one may need the mapping of sO2 within tissue. To achieve this goal hybrid techniques can serve as an optimal solution.
Optoacoustic (OA) imaging [3, 4] is one of such hybrid methods. OA imaging is based on the detection of ultrasonic waves generated in the studied biological tissue due to the absorption of probing laser pulses by optical inhomogeneities. As different chromophores have their own absorption spectra, spectral OA measurements allow for the reconstruction of chromophores concentrations. The main blood components, oxy- and deoxyhemoglobin, are the strongest absorbers in the visible wavelength range, which makes it possible to use optoacoustic method to obtain angiographic images. OA imaging combines benefits of optical and ultrasound imaging resulting in high spatial resolution, contrast and functional imaging. One of the most promising directions of OA imaging application is the 3D mapping of sO2 based on OA spectroscopy data, owing to significant difference in oxy- and deoxyhemoglobin absorption spectra, which makes it possible to use OA for various biomedical applications. However, biological tissues are optically inhomogeneous, and their optical properties are a priori unknown, which requires the development of customized approaches for the extraction of physiological parameters. Current trend consists in the application of machine learning techniques [5, 6] using large sets of synthetic OA data, which requires numerical solutions of both optical and acoustic problems. We report on the employment of machine learning based approaches in pixel-by-pixel 3D reconstruction of sO2.
Usually, quantitative OA imaging research aims to achieve an absolute quantification of the absorption coefficient from measured OA signals. In OA imaging, quantification of involves a solution of two ill-posed inverse problems: acoustic reconstruction yielding images of the OA signal p(r,A) for each measurement wavelength
p(r,A) = rH(r,A) (1) and the estimation of from absorbed energy spectrum H(r, A), which can be written as
H(r,X)=^a(r,X)0(r,X) (2) where &(r,A) is distribution of local light fluence.
In this study we trained machine learning algorithm on in silico data generated through Monte Carlo (MC) simulation of absorbed light distribution in model tissue and than estimated sO2 from in vivo OA data. A custom developed MATLAB-based implementation of MC algorithm [7-9] was employed for the generation of absorption maps characterized by the distribution of the absorbed dose in a tissue-like medium containing blood vessels with varying diameters and depths within morphological range. In this study we consider 2 wavelengths and simplified model of rabbit ear, containing 4 blood vessels with different diameters and location depths. Absorption maps were calculated for different configuration of model tissue (vessel diameters and depths) and also different sO2 level.
Calculated absorption maps for different vessel configurations at 532 and 1064 nm were normalized and converted in pixel-wise dataset with ground truth sO2 and embedding depth labels. 800000 samples were divided on train and test datasets and used for training Gradient Boosting Regressor (GBR) [10]. Raw OA data obtained from OA microscope were reconstructed by delay-and-sum reconstruction [11]. Reconstructed OA data with accounting for the tissue surface were also normalized and converted in pixel-wise dataset for presenting to trained model and estimation of sO2. The data normalization was performed by matching the signal ranges from OA and MC data.
The machine learning algorithm was tested on the set of experimental data acquired on rabbit ear in vivo [12] in the course of temperature stimulation (cooling down to 15 °C followed by heating up to 43 °C) of arteriovenous anastomosis (AVA) resulting in the variation of blood saturation level in vessels. Maximum intensity projection (MIP) OA images obtained at 532 nm are similar in both temperature regimes (Fig. 1a), while in cold regime AVA phenomenon (Fig. 1b) and its interruption (Fig. 1c) are visible on OA images obtained at probing wavelengths 1064 nm. This set of in vivo data was employed for testing the developed algorithm (Fig. 2).
a) jLV C)
\ \
Figure 1: MIP of OA reconstructed images obtained at 532 nm (a) and 1064 nm (b-c) at immersion chamber temperature of 15 °C (a,b) and 43 °C (c). All bars are 1 mm.
Figure 2: MIP of reconstructed saturation maps at immersion chamber temperature of 15 °C (a) and 43 °C
(b).
Results of machine learning-based sO2 reconstruction from MIP OA images visualize AVA appearance and interruption (Fig. 2). It shows clear difference in sO2 values obtained in veins in cold and warm regimes. Veins are connected with arteries at 15 °C, which leads to exchange of oxygenated blood between them (Fig.
2a) and high sO2 values in both arteries and veins, while the interruption of AVA at 43 °C leaves only low oxygenated blood in veins (Fig. 2b).
Thus, the developed algorithm demonstrates high potential for sO2 mapping from MIP OA images without accounting for vessel embedding depth. It has obvious perspectives to be employed for different biomedical applications in quantitative OA imaging.
ACKNOWLEDGEMENTS
The study is supported by supported by Russian Science Foundation (project no. 22-29-00074).
REFERENCES
[1] H.-W. Wang, M.E. Putt, M.J. Emanuele, D.B. Shin, E. Glatstein, AG. Yodh, T.M. Busch, Treatment-induced changes in tumor oxygenation predict photodynamic therapy outcome, Cancer Res. 64(20), 7553, 2004.
[2] L. Xiang, D. Xing, H. Gu, D. Yang, S. Yang, L. Zeng, W.R. Chen, Real-time optoacoustic monitoring of vascular damage during photodynamic therapy treatment of tumor. J. Biomed. Opt. 12(1), 014001, 2007.
[3] A. Taruttis, V. Ntziachristos, Advances in real-time multispectral optoacoustic imaging and its applications, Nat. Photonics 9(4), 219, 2015.
[4] X.L. Dean- Ben, D. Razansky, Optoacoustic imaging of the skin, Exp. Dermatol. 30(11), 1598, 2021.
[5] J, Grohl, M. Schellenberg, K. Dreher, L. Maier-Hein, Deep learning for biomedical photoacoustic imaging: A review, Photoacoustics 22, 100241, 2021.
[6] C. Bench, A. Hauptmann, B. Cox, Toward accurate quantitative photoacoustic imaging: learning vascular blood oxygen saturation in three dimensions, J. Biomed. Opt. 25(8), 085003, 2020.
[7] M. Kirillin, A. Khilov, D. Kurakina, A. Orlova, V. Perekatova, V. Shishkova, A. Malygina, A. Mironycheva, I. Shlivko, S. Gamayunov, I. Turchin, E. Sergeeva, Dual-Wavelength Fluorescence Monitoring of Photodynamic Therapy: From Analytical Models to Clinical Studies, Cancers 13(22), 5807, 2021.
[8] D. Kurakina, M. Kirillin, V. Perekatova, V. Plekhanov, A. Orlova, E. Sergeeva, A. Khilov, A. Nerush, P. Subochev, S. Mallidi, I. Turchin, T. Hasan, owards Bimodal Optical Monitoring of Photodynamic Therapy with Targeted Nanoconstructs: A Phantom Study, Appl. Sci. 9(9), 1918, 2019.
[9] V. Perekatova, D. Kurakina, A. Khilov, M. Kirillin, Combined Monte Carlo and k-Wave Simulations for Reconstruction of Blood Oxygen Saturation in Optoacoustics: A Pilot Study, J. Biomed. Photonics Eng. 8(4), 40511, 2022.
[10] P. Prettenhofer, G. Louppe, Gradient boosted regression trees in scikit-learn, in PyData 2014, 2014.
[11] P. Subochev, F. Spadin, V. Perekatova, A. Khilov, A. Kovalchuk, K. Pavlova, A. Kurnikov, M. Frenz, M. Jaeger, Toward Real-Time Giga-Voxel Optoacoustic/Photoacoustic Microscopy: GPU-Accelerated Fourier Reconstruction with Quasi-3D Implementation, Photonics 9(1), 15, 2022.
[12] V. Perekatova, S. Nemirova, A. Orlova, M. Kirillin, A. Kurnikov, K. Pavlova, A. Khilov, A. Kovalchuk, P. Subochev, Three-dimensional dual-wavelength optoacoustic angiography reveals arteriovenous anastomoses, Las. Phys. Lett. 18(4), 045601, 2021.
