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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Conventional deep learning models for image denoising face challenges like high computational load and energy consumption.
• Existing methods often struggle with effective noise reduction without compromising image quality.

Purpose of the Study:

• To develop an energy-efficient denoising pipeline for medical images.
• To improve the quality of input data for neural networks through preprocessing.
• To reduce computational costs and training time for image denoising.

Main Methods:

• Integrated image enhancement (sharpening kernels) and K-means clustering for image segmentation as preprocessing steps.
• Applied a convolutional autoencoder for denoising after preprocessing.
• Evaluated performance on CT and MRI datasets using PSNR, SSIM, and classification accuracy.

Main Results:

• Significant improvements in Peak Signal-to-Noise Ratio (PSNR) from 21.52 dB to 28.14 dB.
• Enhanced Structural Similarity Index Measure (SSIM) from 0.7619 to 0.8690.
• Reduced training time by approximately 20% and lowered GPU utilization.

Conclusions:

• The proposed energy-efficient denoising pipeline effectively improves medical image quality and reduces computational demands.
• The methodology supports sustainable medical imaging by minimizing radiation exposure and repeat scans.
• The approach is suitable for remote diagnostics and telemedicine in low-resource settings.