Quantitative Evaluation of a Deep Learning-based U-Net Model for Denoising Lung CT Images

This study explores the efficacy of the U-Net architecture for denoising lung CT images that have undergone the introduction of artificially added Gaussian noise. Noise in medical imaging can significantly compromise diagnostic accuracy, prompting the utilization of U-Net to mitigate this challenge. We set the noise standard deviations of 0.03, 0.05, and 0.07, employing metrics such as Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and skewness and kurtosis of patch-based statistics difference across a dataset of 100 CT DICOM files. The results exhibited substantial enhancements in evaluating metrics following denoising, underscoring the model's capability to effectively alleviate noise and enhance diagnostic reliability. Despite positive evaluation metric results, blurring effects led to the loss of anatomical information, demonstrating the necessity for further advancements in denoising techniques. This research provides a foundational framework for the development of robust methodologies aimed at improving the quality of medical imaging.