Comparative analysis of noise estimation methods in computed tomography images: Histogram analysis, L2 norm, SSIM, and CNN-based classification with ResNet50

The quality of medical images is frequently compromised by noise, arising from issues such as data collection, raw data processing , and technology constraints. Determining the primary noise type on computed tomography (CT) scans is essential for choosing effective denoising methods and guaranteeing precise image analysis in medical contexts. This study presents a comprehensive framework for noise classification that employs three independent approaches. The first approach employs histogram-based similarity metrics, such as Correlation, KL Divergence , JS Divergence, KS Distance, and Bhattacharyya Distance , to statistically analyze noise characteristics. The second approach depends on the L 2 norm and the structural similarities index (SSIM) to assess similarities between noisy images and reference noise models, facilitating noise identification. The third approach is deep learning-based classification using ResNet50, which directly learns noise patterns from the collected images to facilitate automated categorization. Experimental results demonstrate that Poisson noise is the predominant noise type in the assessed CT datasets, especially in sharp and medium-kernel images. The findings are consistently corroborated across all three approaches, with occasional evidence of Rician noise in certain instances. The existence of additional noise types, including Gaussian , Speckle, and Salt & Pepper noise, was determined to be negligible according to SSIM and aggregated similarity scores. The findings underscore the framework reliability in precisely identifying noise features, aiding in the selection of appropriate denoising approaches, and ultimately improving medical image quality.