Mitigating Common Sources of Image Degradation Using Deep Learning Techniques in Photoacoustic Tomography

Various deep learning approaches to enhance the quality of photoacoustic tomography (PAT) images are discussed. Numerical phantoms of physiological significance - superposition of solid shapes (ellipse and circle) to understand the extent of distortion of tangential resolution, and vasculature phantom which mimics a cross-sectional view of a blood vessel network in tissue - were used to evaluate the performance of the models. The k-Wave toolbox was used to simulate the forward data. It simulates the propagation of ultrasonic waves generated by laser-induced thermoelastic expansion of biological medium (photoacoustic effect). The simulations accounted for fixed center frequency of the transducer (5 MHz) with varying bandwidth (50-80%) and signal-to-noise ratio (20–40 dB) to mirror practical imaging challenges. Moreover, a reduced number of sensors (50) was used during simulations to create a more resource-efficient setup. Several simulations were performed to obtain the forward data and images were generated using the backprojection algorithm. These were then augmented with techniques like random subsampling, scaling, translation and rotation, creating a dataset of approximately 1500 images to prevent overfitting and improve model robustness. The conventional and Dense-UNET models were trained on the augmented data, and their performance was evaluated based on the numerical values of the extensively used metrics- structural similarity index and peak signal-to-noise ratio. Initial results indicate that Dense-UNET offers superior performance in dealing with complex artifacts than the traditional UNET. Future work will focus on training models on more complex phantoms and integrating experimental data to enhance the current dataset and further increase the robustness of the model. We also aim to extend models specifically tailored for heavyweight or lightweight use, to address various needs.