Automated Segmentation of Multiple Sclerosis Lesions in Magnetic Resonance Imaging

Automated segmentation of multiple sclerosis (MS) lesions in magnetic resonance imaging (MRI) has become central to both clinical research and routine care. Precise delineation of lesion load and evolution informs prognosis, therapeutic decisions and monitoring of disease progression. Traditional manual or semi-automated methods are labour-intensive, subject to inter- and intra-observer variability, and ill suited to large multi-centre studies. In response, computational models have evolved from statistical classifiers and atlas-based techniques to advanced machine learning and deep neural networks. These approaches exploit multi-contrast MR sequences—such as T1-weighted, T2-weighted and fluid-attenuated inversion recovery images—to characterise lesion appearance, shape and spatial context. Contemporary frameworks address challenges of scanner heterogeneity, label noise and limited annotated data through federated learning, attention mechanisms and data-driven label correction. Standardised evaluation metrics and challenge platforms have underpinned rigorous comparison of algorithms, revealing that ensemble and consensus strategies can approach human performance. Ongoing efforts focus on improving detection of small cortical lesions, ensuring robustness across sites and integrating lesion segmentation with broader brain morphometry for comprehensive phenotyping.

Research from Nature Portfolio

Recent studies have employed large-scale challenge infrastructure to benchmark segmentation algorithms in a fully automated pipeline against multi-expert reference annotations. These efforts showed that while deep learning models markedly outperform earlier machine learning methods, they still fall short of expert raters on lesion detection metrics. Fusion of multiple algorithm outputs into a consensus segmentation has been shown to narrow this gap, achieving near-expert delineation scores. In parallel, fully convolutional neural networks trained on a combination of multi-modal images and anatomical labels have demonstrated strong generalisation across scanner platforms. Novel architectures that segment both lesions and surrounding grey-matter structures maintain high agreement with human raters, even when applied to out-of-centre datasets, highlighting the importance of integrated modelling of neuroanatomy for robust clinical applications.

Automated Segmentation of Multiple Sclerosis Lesions in Magnetic Resonance Imaging publication trend

The graph below shows the total number of articles in automated segmentation of multiple sclerosis lesions in magnetic resonance imaging across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional neural network (CNN): A class of deep learning models that applies convolutional filters to image data, capturing spatial hierarchies and patterns.

U-Net architecture: A convolutional network design with encoder–decoder pathways and skip connections, widely used for precise medical image segmentation.

Dice similarity coefficient: A statistical measure of overlap between predicted and reference segmentations, ranging from 0 (no overlap) to 1 (perfect match).

Federated learning: A decentralised training paradigm where models are updated locally on site-specific data and only aggregated parameters are shared, preserving patient privacy.

Attention mechanism: A technique that guides neural networks to focus on the most relevant features or regions in input data, improving segmentation precision.

Fluid-attenuated inversion recovery (FLAIR): An MRI sequence that suppresses cerebrospinal fluid signal, enhancing visibility of white matter lesions in MS.

References

1. Improving multiple sclerosis lesion segmentation across clinical sites: A federated learning approach with noise-resilient training. Artificial Intelligence in Medicine (2024).
2. Boosting multiple sclerosis lesion segmentation through attention mechanism. Computers in Biology and Medicine (2023).
3. Objective Evaluation of Multiple Sclerosis Lesion Segmentation using a Data Management and Processing Infrastructure. Scientific Reports (2018).
4. Multiple sclerosis cortical and WM lesion segmentation at 3T MRI: a deep learning method based on FLAIR and MP2RAGE. NeuroImage Clinical (2020).
5. Simultaneous lesion and brain segmentation in multiple sclerosis using deep neural networks. Scientific Reports (2021).