Automated multiple sclerosis lesion segmentation from 3D-FLAIR MRI using R2AUNet: A deep learning approach with recurrent residual and attention mechanisms

Aref Andishgar, Maziyar Rismani, Zahra Mohammadi, Sina Bazmi, Mehrnaz Fathian, Mahdi Esmaeili Shafaei, Mahsa Tajali, Sahar Khazforoosh, Leila Sajjadi, Elham Goodarzi, Mohammadamin Sookhaklari, Sedighe Hooshmandi, Alireza Torabi Jahromi, Maryam Poursadeghfard,
Automated multiple sclerosis lesion segmentation from 3D-FLAIR MRI using R2AUNet: A deep learning approach with recurrent residual and attention mechanisms,
Multiple Sclerosis and Related Disorders,
Volume 102,
2025,
106620,
ISSN 2211-0348,
https://doi.org/10.1016/j.msard.2025.106620.
(https://www.sciencedirect.com/science/article/pii/S2211034825003621)
Abstract: Background
Multiple sclerosis (MS) is a progressive neurological disorder marked by demyelinating lesions in the central nervous system. While MRI is essential for MS diagnosis, manual lesion segmentation is time-consuming and prone to variability. This study develops an automated deep learning (DL) approach for MS lesion segmentation using 3D-FLAIR MRI.
Method
The R2AUNet DL model, incorporating recurrent residual blocks and attention gates within a 3D U-Net framework, was developed. The dataset included 112 MRI scans from 95 MS patients, collected between 2019 and 2023 at Shiraz Picture Archiving and Communication System (PACS). All patients had a confirmed MS diagnosis based on clinical assessments and the 2017 McDonald criteria, with manual lesion segmentations from expert neurologists as ground truth. The model was trained using an optimized preprocessing pipeline (brain extraction, bias field correction, intensity normalization, and 3D patch extraction). Dice Similarity Coefficient (DSC), specificity, sensitivity, F1-score, and precision were used to evaluate performance.
Result
The proposed model achieved a DSC of 0.768 on the test set, with an accuracy of 0.998, sensitivity of 0.765, and precision of 0.825, demonstrating robust segmentation performance. Qualitative analysis confirmed the model's ability to accurately delineate lesions while reducing false positives.
Conclusion
R2AUNet provides a reliable and automated solution for MS lesion segmentation from 3D-FLAIR MRI, offering high accuracy and efficiency. This study demonstrates that an automated DL model can achieve accurate MS lesion segmentation using only 3D-FLAIR MRI, eliminating the need for multiple MRI sequences. This simplifies data acquisition, reduces computational complexity, and improves accessibility, particularly in resource-limited settings.
Keywords: Medical image processing; Multiple sclerosis; Convolutional neural network; Lesion segmentation; Deep learning