Convolutional Neural Network Improves MS Lesion Segmentation

The developers of a new convolutional neural network say their method can improve accuracy in multiple sclerosis (MS) lesion segmentation using standard diagnostic imaging.

neurons | Image credit: Matthieu - stock.adobe.com

In a new study published in the Journal of Neuroimaging, they say their method outperforms existing segmentation algorithms.1

Focal brain lesions are a key diagnostic marker of MS. They can also be used to assess treatment response and predict future disability, noted corresponding author Francesco La Rosa, PhD, of the Icahn School of Medicine at Mount Sinai Health System, and colleagues. The number and volume of lesions are also “essential” data points in MS research, they added.

“Reliable lesion segmentation provides the foundation for studying advanced imaging biomarkers, including central vein sign and paramagnetic rim lesions,” they wrote.

MS lesions can be identified using T2-weighted (T2w) magnetic resonance imaging (MRI), but the process of manually segmenting MS lesions is problematic because it is time-consuming and subject to interrater variability, the investigators said. New advances in machine learning and artificial intelligence (AI) have made it possible to automate lesion segmentation. A number of different approaches have been developed to accomplish the task, and automated MS segmentation was even the subject of an international competition held last year, in which the Icahn School of Medicine took part.2

La Rosa and colleagues said the most successful of automated segmentation approaches are based on convolutional neural networks. Still, they said the “vast majority” of automated methods require the use of both T2w fluid-attenuated inversion recovery (FLAIR) MRI and a T1-weighted image. Most of the new algorithms have also been trained on research-grade scans that are not always acquired or feasible in clinical practice.

La Rosa and colleagues therefore sought to develop an automated method that could be effective using only T2w FLAIR images. The team created a deep-learning based algorithm using 668 scans from patients with MS. The scans were acquired using 1.5 and 3 T MRI scanners. They named their algorithm FLAMeS, short for FLAIR Lesion Analysis in Multiple Sclerosis.

In the study, the developers compared FLAMeS to 2 publicly available algorithms using both research- and clinical-grade MRI scans. The “benchmark methods” were the Lesion Segmentation Toolbox (LST) methods and the Sequence Adaptive Multimodal SEGmentation (SAMSEG) algorithm. The LST “toolbox” is actually a package of 3 distinct methods; La Rosa and colleagues used one of the methods when only a FLAIR image was available, and they used a second, AI-based method when both FLAIR and T1w scans were available.

The investigators used 3 external datasets to evaluate the algorithms. The performance of the methods was assessed in 2 different manners. Two blinded experts were tasked with qualitatively assessing the algorithms’ performance. In addition, quantitative assessment was performed by comparing automated and ground truth lesion masks using standard segmentation metrics, the authors said.

In the qualitative review of 20 scans, 1 expert chose FLAMeS as the most accurate segmentation method in 15 cases, and the other chose FLAMeS in 17 cases. In the quantitative analysis, FLAMeS also outperformed the other methods, with a higher positive predictive rate and a superior false positive rate, La Rosa and colleagues said. They added that the other 2 methods missed large and small lesions, while FLAMeS only missed lesions smaller than 10 mm3.

The investigators said it is possible that their algorithm outperformed the others because it has superior underlying architecture, or because it had better training data. They said they believe it is likely that both factors played a role.

The investigators said they have publicly released the model in hopes of allowing others to fine-tune the method.

“By offering improved accuracy and resilience to variations in image quality, resolution, and acquisition protocols, FLAMeS represents a valuable tool for MS research applications,” they said.

References:

1. Dereskewicz E, La Rosa F, Dos Santos Silva J, et al. A novel convolutional neural network for automated multiple sclerosis brain lesion segmentation. J Neuroimaging. 2025;35(5):e70085. doi:10.1111/jon.70085
2. Antonacopoulos A, Chaudhuri S, Chellappa R, et al. ICPR 2024 Competition on multiple sclerosis lesion segmentation—methods and results. Pattern Recognit Compet. Published online November 30, 2024. doi:10.1007/978-3-031-80139-6_1