Optimal Hyperspectral Band Selection for Tissue Oxygenation Mapping with Generative Adversarial Network.

Tissue oxygenation assessment using hyperspectral imaging is an emerging technique for the diagnosis and pre- and post-treatment monitoring of ischemic patients. However, the high spectral resolution of hyperspectral imaging leads to large data sizes and a long imaging time. In this study, we propose a method that utilizes multi-objective evolutionary algorithms to determine the optimal hyperspectral band combination when developing a deep learning model for predicting tissue oxygenation from hyperspectral images. Our results confirm that the deep learning model effectively predicts tissue oxygenation images for various oxygenation states. Moreover, we demonstrate that a high-performance prediction model can be developed using only a small number of spectral bands, indicating the potential for more efficient non-contact tissue oxygenation mapping with the proposed method.Clinical Relevance- The proposed method allows for the non-contact and efficient acquisition of two-dimensional tissue oxygenation information in various oxygenation states.

Sinogram-based super-resolution in PET.

Spatial resolution is intrinsically limited in positron emission tomography (PET) systems, mainly due to the crystal width. To increase the spatial resolution for a given crystal width, mechanical movements such as wobble and dichotomic motions are introduced to the PET systems. However, multiple sinograms obtained through such movements provide oversampled data. In this paper, to increase the spatial resolution, we present a novel super-resolution (SR) scheme that employs multiple sinograms. For SR, we first propose a blur kernel estimation scheme through a Monte Carlo simulation. Based on the estimated blur kernel, we adopt a maximum a posteriori expectation maximization method in estimating a high-resolution sinogram from multiple low-resolution sinograms. The proposed algorithm provides noticeable improvement of the spatial resolution in real PET images.

Image resolution improvement based on sinogram super-resolution in PET.

One of the limits of PET imaging is the low spatial resolution due to a predetermined detector width. To overcome this limit, we may increase the number of samples by using the wobbling motion. Since the line spread function (LSF) of the sinogram is determined by the detector width, however, the increase of the number of samples is not sufficient to improve the sinogram resolution. In this paper, based on oversampled data obtained from the wobbling motion, we propose a novel and efficient super-resolution (SR) scheme for the sinogram. Since the proposed SR scheme adopts the penalized expectation maximization (EM) algorithm, it guarantees non-negative values of the super-resolved sinogram data. Through the experiments, we demonstrate that the proposed SR scheme can noticeably improve the spatial image resolution.