[Paper Review] Using a Generative Adversarial Network for CT Normalization and its Impact on Radiomic Features

Leihao Wei, Yannan Lin+1 more|arXiv (Cornell University)|Jan 22, 2020

Radiomics and Machine Learning in Medical Imaging16 references4 citations

TL;DR

This study proposes a 3D generative adversarial network (GAN) to simultaneously normalize low-dose and thick-slice CT scans to standard-dose, thin-slice reference images, improving perceptual image quality and significantly reducing variability in radiomic features. The GAN outperformed a CNN baseline by improving perceptual similarity by 35% and reducing radiomic feature error in nine out of nine features (p < 0.05), demonstrating its potential for standardizing heterogeneous CT protocols in radiomics and CADx systems.

ABSTRACT

Computer-Aided-Diagnosis (CADx) systems assist radiologists with identifying and classifying potentially malignant pulmonary nodules on chest CT scans using morphology and texture-based (radiomic) features. However, radiomic features are sensitive to differences in acquisitions due to variations in dose levels and slice thickness. This study investigates the feasibility of generating a normalized scan from heterogeneous CT scans as input. We obtained projection data from 40 low-dose chest CT scans, simulating acquisitions at 10%, 25% and 50% dose and reconstructing the scans at 1.0mm and 2.0mm slice thickness. A 3D generative adversarial network (GAN) was used to simultaneously normalize reduced dose, thick slice (2.0mm) images to normal dose (100%), thinner slice (1.0mm) images. We evaluated the normalized image quality using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). Our GAN improved perceptual similarity by 35%, compared to a baseline CNN method. Our analysis also shows that the GAN-based approach led to a significantly smaller error (p-value &lt; 0.05) in nine studied radiomic features. These results indicated that GANs could be used to normalize heterogeneous CT images and reduce the variability in radiomic feature values.

Motivation & Objective

• To address variability in radiomic features caused by differences in CT acquisition protocols, such as dose level and slice thickness.
• To develop a unified deep learning method that normalizes both dose and slice thickness simultaneously, rather than treating them separately.
• To improve perceptual image quality in normalized CT scans using a 3D GAN with spectral normalization for training stability.
• To evaluate whether GAN-based normalization reduces error in radiomic features compared to original heterogeneous scans and CNN-based baselines.
• To assess the impact of normalization on radiomic feature consistency across diverse acquisition parameters, supporting robust CADx systems.

Proposed method

• A 3D conditional GAN with spectral normalization was trained to map low-dose (10%, 25%, 50%) and thick-slice (2.0mm) CT scans to standard-dose (100%) and thin-slice (1.0mm) reference scans.
• The generator network learns to reconstruct high-quality, perceptually realistic CT volumes from degraded inputs, while the discriminator distinguishes real reference scans from generated ones.
• The model was trained using a combination of adversarial loss and perceptual loss, with image quality evaluated via PSNR, SSIM, and LPIPS.
• A baseline CNN was trained with mean squared error (MSE) loss for comparison, using the same input-output pairs.
• Radiomic features were extracted from nodules in test scans and compared across original, CNN-generated, and GAN-generated images to quantify error relative to the reference.
• Statistical significance of feature error differences was assessed using paired Wilcoxon signed-rank tests (p < 0.05).

Experimental results

Research questions

• RQ1 Can a 3D GAN effectively normalize both low-dose and thick-slice CT scans to a standard reference protocol in a single unified framework?
• RQ2 Does GAN-based normalization improve perceptual image quality more than a CNN baseline, as measured by LPIPS, PSNR, and SSIM?
• RQ3 To what extent does GAN-based normalization reduce the error in radiomic features compared to original heterogeneous scans and CNN-based outputs?
• RQ4 Are the reductions in radiomic feature error statistically significant across multiple features and normalization scenarios?
• RQ5 How does the GAN-based normalization impact the consistency of texture- and morphology-based features critical for CADx systems?

Key findings

• The GAN improved perceptual similarity by 35% compared to the CNN baseline, as measured by LPIPS, with the largest gain observed in the most extreme normalization scenario (10% dose, 2.0mm slice).
• Despite lower PSNR and SSIM values, the GAN-generated images exhibited superior perceptual quality with sharper edges and richer textures, particularly in nodule margins.
• The GAN-based method achieved significantly lower mean error in nine radiomic features compared to the original heterogeneous scans (p < 0.05), with the largest p-value being 0.035.
• The GAN reduced feature error compared to the CNN baseline in eight out of nine radiomic features, with statistical significance marked by asterisks in the results.
• The normalized images showed reduced variability in radiomic features, indicating improved consistency across different acquisition protocols, which is critical for reliable CADx performance.
• The study demonstrates that joint normalization of dose and slice thickness via a 3D GAN leads to more reliable and consistent radiomic feature extraction than conventional denoising or super-resolution alone.