Comprehensive Study on Denoising of Medical Images Utilizing Neural Network Based Auto-Encoder

TL;DR

This paper investigates how auto-encoder neural networks can effectively denoise spectral medical images, particularly fetal movement images, improving information retrieval despite high noise levels.

Contribution

It introduces a deep learning auto-encoder approach tailored for denoising spectral fetal movement images, demonstrating successful noise mitigation even with substantial Super-Gaussian noise.

Findings

Auto-encoder effectively reduces noise in spectral images.

High correlation noise can be restored with minimal error.

Improved image quality aids fetal motion analysis.

Abstract

Fetal motion discernment utilizing spectral images extracted from accelerometric data incident on pregnant mothers abdomen has gained substantial attention in the state-of-the-art research. It is an essential practice to avoid adverse scenarios such as stillbirths and intrauterine growth restrictions. However, this endeavor of ensuring fetus safety has been arduous due to the existence of random noise in medical images. This novel research is an in depth approach to analyze how the interference of different noise variations affect the retrieval of information in those images. For that, an algorithm employing auto-encoder-based deep learning was modeled and the accuracy of reconstruction of the STFT images mitigating the noise has been measured examining the loss. From the results, it is manifested that even a substantial addition of the Super-Gaussian noises which have a higher correlation of the frequencies possessed by the Fetal movement images can be restored successfully with the slightest error.