Preserving Critical Signals in Magnetic Image Denoising: A Deep Learning Approach with Selective Feature Preservation
Keywords:
MRI Denoising, Deep Learning for Medical Imaging, Residual Learning Framework, Attention Mechanism in Neural Networks, Structural Similarity Index (SSIM), Feature Preservation, Convolutional Neural Networks (CNNs), Supervised Learning, Medical Image Reconstruction, Noise Reduction in MRIAbstract
Noise contamination is a major issue in medical imaging because it affects the clarity of structures and can impact how doctors make diagnoses. To address this, this study introduces a new deep learning method called DenoiseNet. The main goal is to reduce noise without losing important details of the body’s anatomy, which is a challenge with traditional filtering techniques and standard CNN models that often smooth out too much and lose key information. DenoiseNet builds on the U-Net structure by adding spatial attention, channel attention, and residual blocks. These components help the model focus on noisy areas, highlight important features, and ensure that the learning process works smoothly. The model uses residual-attention fusion in the bottleneck, extracts important features in the encoder, and restores clear images in the decoder using skip connections and residual attention blocks. A hybrid loss function that combines MSE and SSIM helps balance pixel accuracy with how realistic the image looks, improving both noise reduction and structure preservation. Hybrid DenoiseNet, incorporating spatial and channel attention along with residual U-Net blocks, achieves a PSNR of 32.27 dB and SSIM of 0.9598. The performance is robust in both our Salt & Pepper noise dataset as well as a semi-synthetic MRI dataset—outperforming both BM3D (31.9 dB, 0.9862) and DnCNN (31.5 dB, 0.8826) under identical test conditions. These qualitative gains are a demonstration of improved noise suppression without loss of structural detail. This approach's strength is its capacity to produce encouraging outcomes even in the early phases of training, exhibiting consistent performance and the possibility of more gains with more time spent training. In comparison to conventional methods, the model gains improved feature representation and better convergence by including attention and residual learning into the U-Net backbone. When taking everything into account, the proposed DenoiseNet demonstrates that merging residual learning with attention mechanisms on a U-Net structure creates a powerful and effective approach for removing noise from medical images. The results show that the model preserves key anatomical details essential for accurate clinical analysis while also effectively reducing noise. These outcomes highlight DenoiseNet's potential as a robust framework that can be further improved and adapted for different types of medical imaging, paving the way for better patient outcomes and more reliable diagnoses.
Downloads
Published
How to Cite
Issue
Section
License
This is an open Access Article published by Research Center of Computing & Biomedical Informatics (RCBI), Lahore, Pakistan under CCBY 4.0 International License