Energy-Efficient Machine Learning Based Denoising Techniques for Sustainable Medical Imaging

Conventional deep learning models have demonstrated denoising potential, but face challenges such as extensive computational load, energy usage, and training time. This study presents an energy-efficient denoising methodology that integrates image enhancement and K-means clustering as preprocessing techniques to improve input quality before applying neural networks. This study proposes an energy-efficient denoising pipeline integrating image enhancement using sharpening kernels and image segmentation through K-means clustering before the application of a convolutional autoencoder. The preprocessing steps enabled the model to identify anatomical boundaries and separate noise-affected regions, thereby improving the input quality and enhancing training convergence. Preprocessing sharpens key image features and distinguishes noise-affected regions, enabling adaptive thresholding and more effective denoising with reduced computational cost. The proposed model was evaluated using publicly available CT and MRI datasets. Performance was assessed through Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and classification accuracy. The results showed that PSNR improved from 21.52 dB to 28.14 dB; SSIM increased from 0.7619 to 0.8690, and validation accuracy also improved. The integrated preprocessing reduced training time by ~20% and lowered GPU utilization, thus supporting reproducibility and deployment in computationally constrained environments. The methodology supports sustainable medical imaging practices by minimizing radiation exposure, reducing repeat scans, and extending the lifespan of older imaging equipment. This pipeline contributes to sustainable medical imaging by minimizing radiation exposure, reducing repeat scans, and extending the lifespan of legacy imaging equipment. It is also suitable for remote diagnostics, enhancing telemedicine workflows in low-resource settings. Additionally, the approach supports remote diagnostics, making it suitable for telemedicine applications in low-resource settings.