SigmaFusionNet: A Residual-Enhanced Hybrid CNN for Accurate Gaussian Noise Prediction in Medical X-ray Images
Nasima Islam Bithi, Md. Jarez Miah, Md. Raihan Mahamud · 2025
Accurate estimation of noise levels in medical images is critical for effective denoising and subsequent diagnostic interpretation. This paper presents SigmaFusionNet, a hybrid convolutional neural network (CNN) model that integrates the complementary strengths of VGG19 and DenseNet121 architectures, enhanced with residual connections, to regress the standard deviation of Gaussian noise from synthetically degraded X-ray images. Trained on grayscale radiographs from the MURA dataset with noise levels uniformly sampled in the range [0.01, 0.20], the proposed model achieved superior performance, with an R2 score of 0.9955, Mean Absolute Error (MAE) of 0.0030, Root Mean Squared Error (RMSE) of 0.0040, and Mean Absolute Percentage Error (MAPE) of 4.60%. These results establish SigmaFusionNet as a robust and precise preprocessing module for adaptive denoising pipelines in clinical imaging workflows.