Transformer-based super-resolution lung CT images improve visualization of multiple diseases.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, Shenzhen Children's Hospital, 7019 Yitian Road, Futian District, Shenzhen, 518038, China.
- Radiology department, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, Hai Ning Rd. 100, Shanghai, 200080, China.
Abstract
To assess the Transformer-based Swin2SR model for super-resolution (SR) enhancement of lung CT images and its clinical potential. Chest CT scans from 303 patients at three hospitals were retrospectively included. Standard 512-matrix images were enhanced to 1024- and 2048-matrix versions (SR-1024, SR-2048). Image noise and signal-to-noise ratio (SNR) for lung tissue, muscle, and background air were quantified. A panel of radiologists rated overall image quality and lesion visibility using a 5-point Likert scale in a multi-reader, multi-case (MRMC)ā analysis. A total of 303 patients (age 67 [interquartile range: 59-75] years, 184 males [60.7%]) were included, diagnosed with 13 major lung diseases. Quantitative analysis showed no significant differences in image noise and SNR among the three image types for lung tissue, muscle tissue, and background air (all pā>ā0.05). SR-1024 and SR-2048 images showed substantial improvements in overall image quality, with 84.5% and 85.1% of patients showing improvements, respectively. SR processing significantly enhanced lesion visibility and image quality for bacterial pneumonia, bronchitis, solid nodules, atelectasis, bronchiectasis, and lung cancer compared to the standard 512-matrix images (all pā<ā0.05). The Swin2SR model, particularly when generating 2048-matrix images, significantly enhances radiologists' subjective ratings of image quality and lesion visibility in lung CT imaging, providing clear benefits for visualizing lung diseases. This study demonstrates that the Swin2SR model significantly improves radiologists' subjective ratings of image quality and lesion visibility in lung CT images, and highlights the potential of deep learning in enhancing medical image interpretation.