# A Frequency-Aware Self-Supervised Framework for MEMS-OCT Denoising

## TL;DR

This paper introduces a new self-supervised framework to reduce noise in MEMS-OCT images, improving image quality and diagnostic accuracy.

## Contribution

The novel FEN2N framework integrates frequency-aware modules for better noise suppression in MEMS-OCT images.

## Key findings

- FEN2N achieves a 2.3 dB PSNR improvement over the N2N baseline.
- FE-RFB modules contribute to a 0.02 SSIM improvement.
- FEN2N outperforms state-of-the-art methods in preserving structural details while reducing speckle noise.

## Abstract

Optical coherence tomography (OCT) is a key biological sensing and imaging tool widely used in biomedical detection, and its images are often degraded by multiplicative speckle noises—especially when micro-electro-mechanical system (MEMS) mirrors are employed in endoscopic OCT imaging, which reduces visual quality and affects the accuracy of subsequent analysis. Traditional denoising algorithms and supervised deep learning approaches have shown some effectiveness, but they are limited by their reliance on paired noisy–clean data and their insufficient modeling of global structural dependencies. To address these issues, this paper proposes a frequency-domain enhanced UNet based on the Neighbor2Neighbor (N2N) framework (FEN2N). The proposed FEN2N integrates wavelet-guided spectral pooling modules (WSPMs) and frequency-domain enhanced receptive field blocks (FE-RFBs). In this work, OCT images are obtained in a self-constructed MEMS-OCT system. Then the FEN2N is applied to the OCT image dataset. Results show that FEN2N achieves a more than 2.3 dB PSNR improvement over the N2N baseline, while the incorporation of FE-RFB contributes to a 0.02 improvement in SSIM. In addition, FEN2N outperforms several state-of-the-art methods, effectively suppressing speckle noise while preserving fine structural details that are important for clinical diagnosis.