An Automatic Unmixing Approach to Detect Tissue Chromophores from Multispectral Photoacoustic Imaging

Abstract

Multispectral photoacoustic imaging has been widely explored as an emerging tool to visualize and quantify tissue chromophores noninvasively. This modality can capture the spectral absorption signature of prominent tissue chromophores, such as oxygenated, deoxygenated hemoglobin, and other biomarkers in the tissue by using spectral unmixing methods. Currently, most of the reported image processing algorithms use standard unmixing procedures, which include user interaction in the form of providing the expected spectral signatures. For translational research with patients, these types of supervised spectral unmixing can be challenging, as the spectral signature of the tissues can differ with respect to the disease condition. Imaging exogenous contrast agents and accessing their biodistribution can also be problematic, as some of the contrast agents are susceptible to change in spectral properties after the tissue interaction. In this work, we investigated the feasibility of an unsupervised spectral unmixing algorithm to detect and extract the tissue chromophores without any a-priori knowledge and user interaction. The algorithm has been optimized for multispectral photoacoustic imaging in the spectral range of 680-900 nm. The performance of the algorithm has been tested on simulated data, tissue-mimicking phantom, and also on the detection of exogenous contrast agents after the intravenous injection in mice. Our finding shows that the proposed automatic, unsupervised spectral unmixing method has great potential to extract and quantify the tissue chromophores, and this can be used in any wavelength range of the multispectral photoacoustic images.

Keywords: blind source separation; optoacoustic; photoacoustic; spectral imaging; unsupervised unmixing.

Figure 1

(a) Schematic representation of the simulated multispectral photoacoustic (PA) images with six inclusions, (b) ideal spectral curves of the inclusions.

Figure 2

(a) Schematic of the tissue-mimicking vessel phantom. (b) The PA absorbance spectral graph of the agents measured by using VevoLab.

Figure 3

(a) Source spectra of oxyhemoglobin, deoxyhemoglobin, and background extracted by the non-negative matrix factorization (NNMF) algorithm. Abundance maps of (b) deoxyhemoglobin, (c) oxyhemoglobin, and (d) background.

Figure 4

(a) Quantitative evaluation of the prominent source components (oxyhemoglobin and deoxyhemoglobin), per each circular region of the synthetic phantom. (b) Overlapped abundance maps of oxyhemoglobin and deoxyhemoglobin.

Figure 5

(a) Spectral absorption curves of the detected source components by NNMF, from 2-D spectral PA images of the tissue-mimicking vessel phantom. The overlapped abundance 2-D maps (b) and 3-D maps (c) of the detected source components: ICG and MB.

Figure 6

Pre-injection conditions: (a) ultrasound (US) image of the kidney–spleen view; respectively (b) photoacoustic (PA) image obtained at 880nm, and (c) $S{O}_2$ map.

Figure 7

Post-injection conditions: (a) spectral signature of the endmembers obtained by using NNMF and abundance distribution maps of (b) deoxyhemoglobin, (c) oxyhemoglobin, and (d) ICG.