Impact of CT Intensity and Contrast Variability on Deep-Learning-Based Lung-Nodule Detection: A Systematic Review of...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Lung cancer is a leading cause of cancer mortality globally.
• Early detection via low-dose computed tomography (LDCT) significantly improves survival.
• Variability in CT acquisition parameters (e.g., Hounsfield Unit calibration, reconstruction kernels) hinders deep learning (DL) system performance.

Purpose of the Study:

• To systematically review preprocessing and harmonization strategies for mitigating CT intensity variability in lung cancer detection.
• To evaluate the effectiveness of different approaches in improving the robustness and generalizability of DL systems.

Main Methods:

• Systematic review following PRISMA 2020 guidelines, searching major academic databases for studies from 2020-2025.
• Inclusion of 100 eligible studies evaluating preprocessing techniques like contrast enhancement, HU-preserving normalization, physics-informed harmonization, and DL-based reconstruction.
• Analysis of strategies' impact on nodule conspicuity, quantitative integrity, and performance degradation across scanners.

Main Results:

• Perceptual methods (e.g., CLAHE) improved sensitivity but distorted HU values.
• HU-preserving methods (e.g., ComBat, physics-informed denoising) were most effective, reducing performance degradation to <8% while maintaining quantitative integrity.
• Transformer-based architectures showed superior robustness (AUC 0.90-0.92) compared to CNNs (AUC 0.85-0.88).

Conclusions:

• Standardized, HU-faithful preprocessing and harmonization-aware AI models are essential for clinical reliability and vendor-agnostic AI in lung cancer screening.
• Multi-center external validation is critical for robust AI development.
• Heterogeneous reporting of acquisition parameters across studies limits comprehensive synthesis.