Distribution-informed and wavelength-flexible data-driven photoacoustic oximetry

TL;DR

This work tackles the challenge of robust, quantitative sO$_2$ estimation from multispectral photoacoustic imaging by introducing a wavelength-flexible LSTM network for spectral unmixing. It demonstrates that training-data selection, guided by Jensen-Shannon divergence, markedly influences accuracy across simulations, phantoms, and in vivo data, and shows that D_JS can predict suitable training datasets for new applications. The approach outperforms linear unmixing and prior data-driven methods, achieving a broader dynamic range and improved alignment with ground-truth or literature references. Together, the flexible LSTM architecture and D_JS-based training-data optimisation offer a promising route toward clinically robust photoacoustic oximetry, while highlighting the need for realistic simulations, 3D modelling, and domain adaptation to bridge the simulation-to-clinical gap.

Abstract

Significance: Photoacoustic imaging (PAI) promises to measure spatially-resolved blood oxygen saturation, but suffers from a lack of accurate and robust spectral unmixing methods to deliver on this promise. Accurate blood oxygenation estimation could have important clinical applications, from cancer detection to quantifying inflammation. Aim: This study addresses the inflexibility of existing data-driven methods for estimating blood oxygenation in PAI by introducing a recurrent neural network architecture. Approach: We created 25 simulated training dataset variations to assess neural network performance. We used a long short-term memory network to implement a wavelength-flexible network architecture and proposed the Jensen-Shannon divergence to predict the most suitable training dataset. Results: The network architecture can handle arbitrary input wavelengths and outperforms linear unmixing and the previously proposed learned spectral decolouring method. Small changes in the training data significantly affect the accuracy of our method, but we find that the Jensen-Shannon divergence correlates with the estimation error and is thus suitable for predicting the most appropriate training datasets for any given application. Conclusions: A flexible data-driven network architecture combined with the Jensen-Shannon Divergence to predict the best training data set provides a promising direction that might enable robust data-driven photoacoustic oximetry for clinical use cases.

Figures (7)

• Figure 1: Overview of the methods used.A Photon transport was simulated with a Monte Carlo light model for each of 25 distinct datasets adapted from a baseline tissue assumption (BASE). Spectra with simulated amplitude for each wavelength were extracted from these simulations. B A deep learning network based on a long short-term memory (LSTM) network was introduced to enable greater flexibility regarding input wavelengths for analysis. The hidden state of the LSTM was passed to fully connected layers, which output the estimated blood oxygenation sO$_2$. C The performance of the LSTM-based method when trained on datasets with different tissue simulation parameters was tested across different datasets, ranging from in silico simulations, in gello phantom measurements, to in vivo measurements. This figure was created with Inkscape using BioRender assets.
• Figure 2: The LSTM-based method shows wavelength flexibility.A LSTMs were trained with varying numbers of wavelengths ($N_\lambda$) to show that with an increasing number of wavelengths, the accuracy of the predictions increases. B LSTM trained at a given $N_\lambda$ can be applied to data with different $N_\lambda$ but yields best results when $N_\lambda$ of the test spectra matches that of the training spectra (indicated by the green vertical line).
• Figure 3: Dimensionality reduction and cross-validation reveal systematic differences between training datasets.A An LSTM-based network trained on each dataset is then applied to every other dataset and all $\epsilon \text{sO}_2$ (median absolute error in percentage points) can be visualised as a performance matrix. Dataset names are shortened for visibility but are detailed in Table \ref{['tab:datasets']}. B Uniform Manifold Approximation and Projection (UMAP) projections of the four representative example datasets onto an embedding of all training data. C Mapping the ground truth sO$_2$ onto the same projection reveals a correlation along the first UMAP axis.
• Figure 4: The Jensen-Shannon divergence (D$_\text{JS}$) can predict estimation performance.A D$_\text{JS}$ correlates with the median absolute sO$_2$ estimation error $\epsilon \text{sO}_2$ when applying all networks, each trained on a distinct training dataset, to the baseline dataset (BASE). B D$_\text{JS}$ for the D$_2$O flow phantom data, which shows a similar correlation with $\epsilon \text{sO}_2$. C After removing two outliers, D$_\text{JS}$ shows the same degree of correlation with $\epsilon \text{sO}_2$ for the H$_2$O flow phantom data.
• Figure 5: Estimation of flow phantom data highlights performance dependence on training dataset. Three example images of the D$_2$O flow phantom are shown at different time points (0, 22, 44 min) displaying A the photoacoustic signal intensity at 700 nm with a red contour marking the blood-carrying tube and B the sO$_2$ estimations from different methods. We visually compare the performance of linear unmixing (LU, C), learned spectral decoloring (LSD, D) and the lstm-based method (LSTM, E) by plotting the sO$_2$ estimations over the same image section and time points shown in A.