Clinical CT image super-resolution using pixel-wise hybrid high-dimension mapping-based implicit neural representation

High-resolution computed tomography (HRCT), with its precise fine structure imaging character, is of growing importance in modern clinical practice. Emerging deep-learning-based CT super-resolution (SR) approaches, particularly those arbitrary-scale SR networks employing implicit neural representation (INR), are demonstrated with promising results. Nevertheless, existing INR algorithms are still hindered by the inferior ability to capture CT fine structural detail since the multilayer perceptron module (MLP) in INR is biased to learn high-frequency components. In this paper, we propose to incorporate a pixel-wise hybrid high-dimension mapping (HHM) module into INR to alleviate the above issues. The HHM module applies sinusoidal function simultaneously on both latent features and spatial coordinates before MLP, projecting the incorporated spatial and image features into a higher dimensional space, to force the INR network to learn more high-frequency details. The fine structural detail in arbitrary-scale SR CT is thus enhanced. We train the proposed network using real low- and high- resolution clinical CT images rather than using down-sampling images, which is more practical in the clinic. We validate the proposed method in detail using qualitative and quantitative evaluations on the thoracic and pelvic dataset, and the results indicate that our method is accurate and robust, outperforming the other deep learning methods.