Performance evaluation of medical image denoising using convolutional autoencoders

• Main Authors: Ismail, Amelia Ritahani, Azhary, Muhammad Zulhazmi Rafiqi, Noor Azwan, Nor Aiman Zaharin, Ismail, Ahsiah, Alsaiari, Nisrin Amer
• Format: Proceeding Paper
• Language: English
• Published: 2024
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://irep.iium.edu.my/116733/1/Performance_Evaluation_of_Medical_Image_Denoising_using_Convolutional_Autoencoders%203.pdf; http://irep.iium.edu.my/116733/

Convolutional Autoencoders (CAEs) are neural net- work architectures specifically designed for image-processing tasks. CAEs also show a great performance in image denoising as it has the capability to efficiently eliminate noise from images while retaining crucial features and structural integrity. However, the effectiveness of CAEs in denoising tasks can be limited by the specific characteristics and complexities present in the images; for example, different intensities and distribution of noise will lead to challenges in achieving consistent and high- quality denoising results across diverse datasets like medical images. This paper, therefore, explores the capabilities of CAEs in effectively denoising images corrupted by varying levels of Gaussian noise. The CAE’s performance is evaluated under different levels of noise severity using two datasets: PathMNIST and Brain MRI Scan images. Findings indicate that at a low noise level, CAEs achieve impressive results with PathMNIST images, demonstrating high image fidelity. However, Brain MRI Scan images show moderate denoising suggesting opportunities for improvement. As noise severity increases, both datasets experience reduced denoising performance, highlighting the need for enhanced CAE configurations to maintain image quality in critical applications like medical diagnosis.