Generalization of an optoacoustic model-based image reconstruction method for linear probes through deep learning

Multispectral optoacoustic tomography (MSOT) is an emerging imaging technique that combines the advantages of both optical and ultrasound imaging to achieve high-resolution functional imaging. It offers several key advantages for imaging biological tissues and organisms, making it a promising tool for various applications. It’s based on the photoacoustic effect that is induced when a nanosecond laser pulse illuminates the tissue. The light energy at specific wavelengths is absorbed by chromophores, that can be endogenous (like haemoglobin, lipids, etc) or exogenous, which leads to a thermoelastic expansion resulting in the generation of ultrasound waves. These waves are detected by the probe and can then be processed by an elaboration system in order to obtain the raw signal, known as synogram. This thesis investigates various commonly employed phtoacoustic image reconstruction algorithms. Traditional methods such as backprojection reconstruction and iterative model based are analyzed. Backprojection is very efficient in terms of computational time, this makes it useful for obtaining real-time reconstructions. However, it suffers from low spatial resolution and is more subject to artifacts than iterative reconstruction methods. Model based relies on accurate modeling of the imaging process, including knowledge of the system, material properties, and other factors. In this way it fills the disadvantage related to the low quality of reconstructions but it’s a limited technique due to the high computational time. To try to combine the advantages of these two techniques, several neural network-based algorithms have been proposed in literature. In this thesis work, a deep learning-based approach for reconstructing optoacoustic images called DeepMB proposed by iThera Medical GmbH is presented. The dataset needed for network training was obtained through an opensource toolbox that allows synthetic data to be obtained using real world images. The toolbox takes the images appropriately scaled and converted to grayscale as input and through forward-model makes it possible to obtain the synograms as if they had been acquired by the probe. It also makes it possible to reconstruct images from the synthetic synograms by backprojection and iterative model-based by providing the generation of models for different speeds of sound. The toolbox was provided to be used with iThera's Acuity Echo convex geometry probe. A central part of my work was to extend the code for applying the toolbox also for the L11-5v Verasonics linear probe. The Opotek Phocus Mobile SE laser and the Verasonics Vantage-256 system presented at the Polytechnic of Turin laboratory were employed to acquire agar phantom images of lead inserts. The network used has a U-Net structure, this takes synograms as input returning model-based like reconstructed images as output. An important factor that enhance the performance of the network is the possibility of including speed of sound as an additional layer to the network using one-hot-encoding and turning it into a trainable parameter. DeepMB represents a method that can be applied in real-time and allows reconstructions to be obtained with quality on par with model-based.