Clinical applications of quantitative photoacoustic imaging

Abstract

Photoacoustic imaging is a medical imaging tool increasingly used in clinical research. It enables the imaging of molecules, including melanin, haemoglobin, lipids, collagen and water deep inside the body, with high resolution and high sensitivity. This molecular information has made it an appealing tool for detecting and monitoring a range of diseases, from degenerative muscular disorders to cancer. Here, I take three steps towards the clinical translation of photoacoustics: tackling skin colour bias, applying it to a new disease and improving the transparency of computational methods. To characterise the effects of skin colour on photoacoustic imaging, I carried out a technical study using simulations, tissue-mimicking phantoms, and pigmented mice. Linear unmixing and a machine learning approach were compared to estimate blood oxygenation (sO2). A consistent trend of increasing sO2 with increasing skin pigmentation was observed in all three settings, suggesting that spectral colouring was leading to decreased light penetration at shorter wavelengths compared to longer wavelengths. The results, however, hinted that further effects were at play. While the machine learning unmixing approach worked well for the simulations and phantoms, it did not work well in the mice due to substantial regions of negative pixels in the pigmented mice. To explore this phenomenon further, I carried out a healthy volunteer study. Data from a diverse cohort of subjects was acquired, and photoacoustic measurements of several major blood vessels and muscles were made. Blood oxygenation was then estimated using linear unmixing, revealing a similar trend to that observed in models. However, rather than seeing a consistent increase in sO2 with skin pigmentation, the estimated sO2 appears to reach a peak and then decrease again beyond a specific limit. I proposed a possible mechanism: the reflection of photoacoustic waves originating in the epidermis. I used simulations to explore this possibility and revealed that the reflected epidermis signal can be detected in the measured photoacoustic time series and may contribute to this bias. If true, this poses a severe problem to existing correction methods, as most are based purely on the optical component of the photoacoustic problem and neglect acoustic artefacts. Notwithstanding the challenges of applying photoacoustics equitably in people with different skin colours, the technology has been successfully applied previously to various disease types. Here, I analysed photoacoustic data from a clinical study of patients with mitochondrial disease. The patients were given 60 % oxygen to assess its effects. Photoacoustic imaging and pulse oximetry were conducted at three timepoints: before, during, and after oxygen delivery. I showed that photoacoustic imaging is sensitive to changes in the blood oxygenation induced by the changing breathing gas and that characteristic multispectral differences between affected patients and healthy volunteers may be associated with lipid content. To improve transparency in photoacoustic imaging, I developed a Python data analysis toolkit called PATATO. It provides open-source implementations of commonly used image reconstruction algorithms and a convenient interface for technical and non-technical users. The parameters of the reconstruction algorithm were optimised, showing that model-based image reconstruction improves the image quality and the photoacoustic spectra derived from clinical data. Overall, I showed that photoacoustic imaging is a promising clinical tool but shows a clear bias due to melanin in the epidermis. Transparency, openness, and improved reporting will all be required to ensure equity in photoacoustic clinical trials going forward. More advanced hardware, software and standardisation protocols will be required to eliminate skin colour bias.