Abstract
Accurate and reliable medical image segmentation is essential for computer-aided diagnosis and formulating appropriate treatment plans. However, noise often significantly reduces diagnostic accuracy and complicates treatment planning. Therefore, noise reduction in medical imaging is paramount, as it not only improves diagnostic accuracy but also contributes to enhanced treatment efficacy and minimizes patient risk. Prior methods have explored frequency-domain approaches to accelerate convolutional operations or combine frequency-based features with spatial convolutions. However, most only partially integrate Fourier-based processing and thus fail to fully exploit its advantages. We propose a novel neural architecture, FFTMed, that operates directly in the frequency domain, harnessing its resilience to noise and uneven brightness while also reducing computational overhead. Notably, FFTMed requires no additional noise augmentation during training yet remains resilient when confronted with noisy test images, demonstrating its effectiveness in real-world medical image segmentation tasks. Additionally, we propose a new benchmark incorporating various levels of noise to assess susceptibility to noise attacks. The experimental results demonstrate that FFTMed not only effectively eliminates noise and consistently achieves accurate image segmentation but also shows robust resistance to imperceptible adversarial attacks compared to other baseline models. The datasets generated and analysed during this study have been deposited in the Zenodo repository and are openly accessible at https://zenodo.org/records/15310397 . The source code to reproduce all experiments is publicly available at https://github.com/HySonLab/LightMed .