Clinical CT Image Super-Resolution Using Pixel-Wise Hybrid High-Dimension Mapping-Based Implicit Neural Representation
Abstract
Purpose
High-resolution computed tomography (HRCT) 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. This work aims to address this limitation by enhancing the ability of INR networks to reconstruct high-frequency details in arbitrary-scale CT SR.
Methods
We propose a pixel-wise hybrid high-dimension mapping (HHM) module integrated into the INR framework. The HHM module applies sinusoidal functions to both latent features and spatial coordinates before the MLP, projecting the combined spatial and image features into a higher-dimensional space. This forces the network to focus more on learning high-frequency details. The model was trained using real clinical CT images (low- and high-resolution pairs) rather than synthetically downsampled images, ensuring practical relevance in clinical settings.
Results
Qualitative and quantitative evaluations on thoracic and pelvic CT datasets demonstrated that the proposed method achieves superior accuracy and robustness compared to other deep learning-based SR approaches. Compared to conventional INR methods, the proposed approach with HHM module improved the Peak Signal-to-Noise Ratio (PSNR) of SRCT images from 38.03 dB to 40.16 dB and increased the Structural Similarity Index Measure (SSIM) from 0.9391 to 0.9580. The enhanced high-frequency detail reconstruction significantly improved the quality of arbitrary-scale super-resolution results.
Conclusion
The proposed HHM module effectively addresses the limitations of existing INR-based CT SR methods, enabling more precise fine-structure imaging. The use of real clinical data for training further validates its practicality, making it a promising tool for clinical applications requiring high-resolution CT imaging.