Medical images, particularly CT scans, often suffer from noise due to data acquisition processes, raw data handling, and hardware limitations. Accurately identifying the dominant noise type in CT images is crucial for selecting targeted denoising techniques and ensuring reliable analysis in clinical settings. This study introduces an integrated framework for noise classification using three complementary methods. The first leverages histogram similarity metrics—including correlation, KL divergence, JS divergence, KS distance, and Bhattacharyya distance—to statistically profile noise properties. The second relies on L2 norm and structural similarity index (SSIM) to compare noisy images against reference noise templates for precise identification. The third employs deep learning with ResNet50 to automatically detect and categorize noise patterns from raw image data. Experiments reveal Poisson noise as the primary type across evaluated CT datasets, especially in sharp- and medium-kernel reconstructions. All three methods consistently validate this, with minor instances of Rician noise detected occasionally. Other noise varieties, such as Gaussian, speckle, or salt-and-pepper, proved negligible based on SSIM values and overall similarity scores. These results highlight the framework’s robustness in pinpointing noise characteristics, guiding effective denoising strategies, and enhancing medical image quality.
Read the articleAuthors: Mahmoud Nasr, Krzysztof Brzostowski, Rafał Obuchowicz, Fathi E. Abd El-Samie, Adam Piórkowski
Keywords: Image noise, Image processing, Correlation Coefficient, Kullback-Leibler (KL) Divergence, Jensen-Shannon (JS) Divergence, Kolmogorov-Smirnov (KS) Distance, Bhattacharyya Distance, SSIM, L2 norm analysis, Combined metric analysis, CNN, ResNet50