SCAF-DG: A Multi-site Medical Image Denoising With A Domain-Generalized Spatial-Channel Attention Fusion

Images from multiple medical sites usually contain varying noise levels that can affect the generalization performance of the denoising models. Domain generalization (DG), which seeks to learn a model that can generalize to an unseen test domain, has not been extensively explored in medical image de...

Full description

Saved in:
Bibliographic Details
Published in2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) pp. 333 - 340
Main Authors Fiasam, Linda Delali, Rao, Yunbo, Sey, Collins, Klugah-Brown, Benjamin, Tettey, Obed Nartey, Aggrey, Esther Stacy E.B., Agyemang, Isaac Osei, Adjei-Mensah, Isaac, Yang, Yang
Format Conference Proceeding
LanguageEnglish
Published IEEE 18.08.2023
Subjects
attention module
Brain modeling
Domain Generalization (DG)
Feature extraction
image denoising
magnetic resonance images
Magnetic resonance imaging
multi-site images
Noise measurement
Noise reduction
Pattern recognition
PSNR
residual learning
Online AccessGet full text

Cover

More Information
Summary:Images from multiple medical sites usually contain varying noise levels that can affect the generalization performance of the denoising models. Domain generalization (DG), which seeks to learn a model that can generalize to an unseen test domain, has not been extensively explored in medical image denoising for multi-site images. We present an approach for multi-site medical image denoising, which addresses the issue of varying noise levels across different imaging sites. Our proposed method, called SCAF-DG (Spatial-Channel Attention Fusion for Domain Generalization), employs two attention units to extract both domain-specific channel features and mutually-invariant spatial features from the noisy images. These features are then combined to form a rich feature space. Additionally, we utilized skip connections to preserve structural information during the denoising process. We evaluate SCAF-DG on publicly available brain magnetic resonance image (MRI) datasets from three different imaging sites, and demonstrate an average increase of 1.01 in peak signal-to-noise ratio (PSNR) and 0.016 in structural similarity index measure (SSIM) for denoising images. Our results demonstrate that SCAF-DG has the potential to generalize well to new, unseen test domains, and outperforms existing state-of-the-art denoising models on multi-site medical images.
DOI:10.1109/PRAI59366.2023.10332061