Scholarlink

<p>Enhancing image quality can be impeded by several factors, including noise, low contrast, artifacts, and limitations imposed by various imaging modalities. Improving the quality of medical images is essential for accurate diagnoses and effective treatment decisions. The analysis of medical images is critical for proper diagnosis. This study investigates autoencoders and U-Net as two of the most powerful deep learning algorithms for image denoising. These algorithms will be combined and evaluated using standard metrics such as the Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Data Structural Similarity Index (DSSIM), and Visual Information Fidelity (VIF). Additionally, denoising operations often employ loss functions like mean-squared error (MSE) and mean absolute error (MAE) during model training to enhance medical images. By leveraging these standard image quality metrics, we were able to develop a new denoising algorithm named hybrid autoencoders and U-Net (HANsU-Net). This novel algorithm is based on a hybrid approach that integrates the autoencoder and U-Net architectures. By leveraging the strengths of both autoencoders and U-Net, HANsU-Net surpasses the performance of each individual algorithm in improving image quality. Since the proposed model is deep and complex, it requires massive computational power that high-performance computing (HPC) systems or supercomputers provide. The quality of data and image structure has been assessed before and after processing using three datasets. Experimental results show that HANsU-Net gives the best results at epoch 70 compared with U-Net or autoencoders. Applying the proposed model, HANsU-Net, to the three chosen datasets has resulted in a better image quality using the above-mentioned metrics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep-learning-based medical image quality amelioration

  • 期刊论文

  • 发表日期：

    2025年10月6日

查看全文
  • Raed Abdullah Althabeti,
  • Ebeid Ali Ebied,
  • Hany Maher Sayed Lala,
  • Kamal Abdelraouf Eldahshan

摘要

Enhancing image quality can be impeded by several factors, including noise, low contrast, artifacts, and limitations imposed by various imaging modalities. Improving the quality of medical images is essential for accurate diagnoses and effective treatment decisions. The analysis of medical images is critical for proper diagnosis. This study investigates autoencoders and U-Net as two of the most powerful deep learning algorithms for image denoising. These algorithms will be combined and evaluated using standard metrics such as the Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Data Structural Similarity Index (DSSIM), and Visual Information Fidelity (VIF). Additionally, denoising operations often employ loss functions like mean-squared error (MSE) and mean absolute error (MAE) during model training to enhance medical images. By leveraging these standard image quality metrics, we were able to develop a new denoising algorithm named hybrid autoencoders and U-Net (HANsU-Net). This novel algorithm is based on a hybrid approach that integrates the autoencoder and U-Net architectures. By leveraging the strengths of both autoencoders and U-Net, HANsU-Net surpasses the performance of each individual algorithm in improving image quality. Since the proposed model is deep and complex, it requires massive computational power that high-performance computing (HPC) systems or supercomputers provide. The quality of data and image structure has been assessed before and after processing using three datasets. Experimental results show that HANsU-Net gives the best results at epoch 70 compared with U-Net or autoencoders. Applying the proposed model, HANsU-Net, to the three chosen datasets has resulted in a better image quality using the above-mentioned metrics.