Enhancing Super-Resolution Network Efficacy in CT Imaging: Cost-Effective Simulation of Training Data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning (DL) super-resolution (SR) models can enhance low-resolution CT images.
• Acquiring adequate training data for CT SR models is a significant challenge.
• Existing methods for simulating thick-slice CT images lack realism or require complex reconstruction.

Purpose of the Study:

• To introduce a simple, realistic method for generating thick CT images from thin-slice CT images.
• To facilitate the creation of high-quality training pairs for DL-based CT SR algorithms.
• To address the data scarcity issue in developing effective CT SR models.

Main Methods:

• A novel simulation technique to generate thick CT images from existing thin-slice CT data.
• Creation of paired training datasets using the proposed simulation method.
• Validation of generated data realism and utility in DL SR model training.

Main Results:

• The generated training pairs closely mimic real data distributions (PSNR = 49.74 vs. 40.66, p < 0.05).
• Radiomics features from CT images generated by the method showed significant correlation with mortality in lung fibrosis patients (HR = 1.19, p < 0.005).
• The proposed method enhances the efficacy and applicability of CT SR models.

Conclusions:

• This study presents the first method to effectively generate paired training data for DL-based CT SR models.
• The developed technique overcomes limitations of previous simulation approaches.
• This work improves the real-world applicability of CT super-resolution in medical imaging.