Comparison of conventional 512-matrix CT images with Swin2SR-based 2048-matrix images in the visualization and diagnosis of lung nodules.

[PURPOSE] Clear visualization and diagnosis of lung nodules depend on the spatial resolution of CT images. Transformer-based generative neural networks can generate super-resolution images. To compare the diagnostic value of standard CT images of 512 × 512 pixels and Swin2SR-based super-resolution images of 2048 × 2048 pixels for lung cancer.

[MATERIALS AND METHODS] The transformer-based Swin2SR model, which can upscale standard 512 × 512 pixel CT images to 2048 × 2048 super-resolution images, was validated with four retrospective datasets, three of which were patient data at three hospitals from January 2018 to December 2020, and another was the public Non-Small Cell Lung Cancer-Radiogenomics dataset. Lung nodules < 3 cm were included to validate the image quality of super-resolution images, and to compare the ability of standard and super-resolution images to display malignancy-associated imaging features and to diagnose lung cancer.

[RESULTS] 1161 nodules (663 malignant and 498 non-malignant) in 1161 subjects (age 60.2 ± 9.9 years, 653 males [56.2%]) were included. Swin2SR-based super-resolution images of these nodules had higher image scores (image quality, sharpness and noise) than standard images (p < 0.001). Among the malignancy-associated imaging features, the Swin2SR-based super-resolution images showed significantly more bubble-like lucency and air bronchogram than standard images (p < 0.001). Of the 663 histologically confirmed malignant nodules, 577 (87.0%) were considered malignant on Swin2SR-based super-resolution images, which was significantly higher than the 529 (79.8%) nodules on standard images (P = 0.037). The accuracy of Swin2SR-based super-resolution images and standard images in diagnosing lung cancer was 0.83 (95% CI 0.81-0.85) and 0.79 (0.76-0.81), respectively.

[CONCLUSION] Swin2SR-based super-resolution 2048 × 2048 pixel CT images can clearly show the malignancy-associated imaging features of lung nodules and improve the diagnostic of lung cancer.