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

Summary

This study proposes an energy-efficient denoising methodology that integrates image preprocessing to improve medical image quality, reduce computational cost, and support sustainable diagnostic practices. The method enhances clarity in low-dose and legacy scans, enabling remote diagnosis while reducing radiation exposure, energy use, and electronic waste.

Abstract

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.

Introduction

Medical imaging plays a pivotal role in diagnostics and treatment planning by offering non-invasive insights into internal anatomical and physiological conditions. Several imaging modalities, X-ray, Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Ultrasound, and Positron Emission Tomography (PET), are routinely used in clinical settings to detect abnormalities, monitor disease progression, and guide interventions^1,2,3. Each modality exhibits unique advantages but is vulnerable to various forms of image degradation caused by instrumentation limitations, acquisition....

Protocol

This study exclusively utilized publicly available, de-identified CT and MRI imaging datasets. No live human or animal subjects were involved. Therefore, no institutional review board (IRB) or ethics committee approval was required.

Method overview
This protocol presents a reproducible pipeline for energy-efficient medical image denoising. It combines preprocessing techniques, including sharpening filters and K-means clustering, with a convolutional neural network (CNN)-based autoencoder to denoise images. This integrated method enhances image quality while reducing training time and hardware energy consumption, support....

Results

Preprocessing and segmentation results
The initial preprocessing phase improved the visibility of critical anatomical boundaries while reducing background interference. As visualized in Figure 7, sharpened images exhibited clearer edge definition, which aided in downstream segmentation. The segmented images created using K-means clustering with values of K = 3 and 5 successfully isolated noise-heavy pixels from diagnostically relevant areas³³. Thi.......

Discussion

This work presents a hybrid denoising approach that integrates image preprocessing with a convolutional autoencoder to enhance diagnostic image quality while optimizing energy usage and computational performance.

The method combines sharpening filters and K-means clustering in the preprocessing phase to improve edge clarity and reduce irrelevant noise, which is then followed by a CNN-based autoencoder for adaptive denoising. This hybrid pipeline reduces unnecessary computational operations and.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Google Colaboratory	Google	N/A	Cloud platform used for model training and testing
Keras (v2.x)	Open-source	N/A	High-level API for TensorFlow used for neural network implementation
Matplotlib (v3.4 or above)	Open-source	N/A	Used for visualization of images and results
Microsoft Excel 365	Microsoft	N/A	Used for result tabulation and analysis
NumPy (v1.21 or above)	Open-source	N/A	Used for matrix operations and numerical computing
NVIDIA Tesla T4 GPU	NVIDIA	N/A	GPU used for accelerated training and inference
Publicly Available Medical Imaging Dataset (CT and MRI Images)	Open Source Databases	N/A	Used as source data for model training, validation, and testing
Python (v3.8 or above)	Python Software Foundation	N/A	Programming language used for model implementation
Scikit-learn (v0.24 or above)	Open-source	N/A	Used for K-means clustering and preprocessing
TensorFlow (v2.x)	table	N/A	Deep learning library used for CNN model development