Enhancing Diagnostic Precision: A Distribution-Based Compressed Denoising Scheme using Transfer Learning for Noise Reduction in Medical Imaging

Accurate interpretation of medical images is vital for diagnosis and treatment planning. However, noise in medical imaging significantly hampers this process, leading to potential diagnostic errors. Traditional denoising techniques often require large datasets or compromise diagnostic details.

This study proposes a Distribution-based Compressed Denoising Scheme (DCDS) leveraging transfer learning to enhance diagnostic precision. The framework analyzes pixel distributions to distinguish normal and noisy pixels using incremental and decremental distribution verification. A chest CT scan dataset with Gaussian-like noise was used to train the model, employing correlation mapping between training and target images for effective variance estimation.

Experimental analysis showed that DCDS significantly improves noise reduction, variance detection, and diagnostic precision. The peak signal-tonoise ratio improved across all evaluated ranges (24-32 dB), with a noise reduction rate exceeding 82% in optimal conditions. DCDS reduced mean error and analysis time, enhancing overall system efficiency. Transfer learning allowed for effective pixel classification and noise prediction with fewer data and computational resources.

Compared to existing denoising methods, DCDS offers better generalization through statistical modeling and knowledge-sharing between training and learning states. However, current validation is limited to CT imaging with Gaussian noise. Broader modality testing and integration of perceptual and clinical grading metrics are needed.

The DCDS framework presents a promising, lightweight, adaptive, and accurate solution for medical image denoising. It holds potential for realtime diagnostic support, but further optimization and clinical validation are necessary to ensure generalizability across diverse imaging modalities and noise types.