An AI-Based Low-Risk Lung Health Image Visualization Framework Using LR-ULDCT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-resolution ultra-low-dose CT (LR-ULDCT) offers reduced radiation exposure but often lacks diagnostic detail compared to high-resolution CT (HRCT).
• Accurate visualization of lung structures and pathologies like ground glass opacity (GGO) is crucial for early disease detection and monitoring.

Purpose of the Study:

• To develop an AI-based visualization framework for low-risk lung health monitoring using LR-ULDCT.
• To achieve diagnostic image quality comparable to HRCT from significantly lower radiation doses (<0.3 mSv).

Main Methods:

• A novel deep cascade network was developed, comprising unsupervised restoration, generative adversarial network (GAN)-based super-resolution (SR), and segmentation.
• The network processes degraded LR-ULDCT to produce restored, super-resolved (SR-ULDCT), and segmented images, including lobe-wise colorization.
• The system was evaluated on real datasets including COVID-19, pneumonia, and pulmonary edema, with comparisons to state-of-the-art methods and verification by radiologists.

Main Results:

• The AI framework successfully enhanced LR-ULDCT images, achieving diagnostic visualization capabilities on par with HRCT.
• The deep cascade network effectively performed restoration, super-resolution, and segmentation, enabling accurate identification and visualization of lung lobes and GGO.
• Case studies demonstrated the system's efficacy on various lung conditions, supported by positive feedback from experienced radiologists.

Conclusions:

• The proposed AI-based framework provides a low-risk, affordable solution for lung health monitoring using LR-ULDCT.
• The system significantly improves the diagnostic power of low-dose CT scans, enabling detailed visualization and analysis of lung pathologies.
• This technology holds promise for widespread clinical application in early lung disease detection and patient management.