Markus Haltmeier,Daniel Obmann,Karoline Felbermayer,Florian Hinterleitner,Peter Burgholzer
Abstract:We investigate resolution in photoacoustic tomography (PAT). Using Shannon theory, we investigate the theoretical resolution limit of sparse view PAT theoretically, and empirically demonstrate that all reconstruction methods used exceed this limit.
What problem does this paper attempt to address?
This paper mainly discusses the sampling and resolution problems in Photoacoustic Tomography (PAT), especially the challenges of image reconstruction from undersampled data. The author analyzed the theoretical resolution limit of PAT under the constraint of maximum bandwidth Ω using Shannon's theory, and experiments showed that the actual reconstruction methods used exceed this limit.
In PAT, the resolution is affected by multiple factors, including acoustic attenuation, limited bandwidth of the detection system, and the number of available data samples. The study focused on two-dimensional PAT with angle undersampling using a circular array of sensors. The paper proposed sampling conditions in the time and angle directions and pointed out that undersampling in the angle direction can result in the inability to recover objects with specific bandwidth Ω, introducing sampling artifacts.
Despite angle undersampling, the study compared several different reconstruction methods through numerical simulations, including standard quadratic Tikhonov regularization, joint ℓ1 regularization combining sparsity and regularity, and deep learning-based methods. The results showed that all methods were able to resolve objects with grid structures to a certain degree.
In the theoretical part, the paper considered the circular geometry model of 2D PAT, discussed the relationship between spatial resolution and Point Spread Function (PSF), and provided conditions for time, spatial, and angular sampling. For an ideal low-pass filter, Ω accurately characterized the spatial resolution induced by the filter.
In the experimental part, the author used a full-optical PAT projection imaging system to collect data and compared different reconstruction methods, including Tikhonov regularization, joint ℓ1 minimization, learned primal-dual iteration, and trained Unet network. The experimental results showed that all tested reconstruction methods were able to recover grid-like patterns from angle undersampled data, with the machine learning method producing the best quality.
The discussion section pointed out that although all methods were successful in recovering grid structures, further research is needed to investigate the resolution performance when reducing angle sampling and changing grid positions, especially to evaluate whether learning-based reconstruction methods can consistently and reliably improve resolution.