Abstract
The acquisition of high-resolution medical images is essential for
the accurate diagnosis and effective treatment of many diseases.
However, obtaining high-resolution images can be limited by factors
such as device limitations and patient exposure to radiation. To
solve this problem, this study proposes the use of super-resolution
techniques based on deep learning to improve the resolution of
computerized tomography images without increasing the patient’s
exposure to radiation. The LoDoPaB-CT image dataset was used.
Five deep learning-based super-resolution techniques - SRCNN,
ESRGAN, SwinIR, HAT and DAT - and the traditional FBP method
were compared. The evaluation metrics included PSNR, SSIM, LPIPS,
NIQE, NRQM and PI. In the tests carried out, the ESRGAN model obtained
the best results, outperforming the other techniques in metrics
such as SSIM, NIQE and PI. On the other hand, the FBP method
showed comparable performance in PSNR, LPIPS and NRQM. These
findings underscore the need to fine-tune the models and highlight
the potential benefit of involving experts in the subjective analysis
process to obtain the best results.