MS Lesion Segmentation Gets Smarter With a New Neural Network
But here’s where it gets controversial: automated, AI-driven MS lesion analysis could redefine how researchers track disease and guide therapy, and not everyone agrees on the pace or the approach.
A team of researchers developed a convolutional neural network that aims to boost the accuracy of MS lesion segmentation using standard diagnostic MRI, potentially improving both research outcomes and patient care.
In a study published in the Journal of Neuroimaging, the researchers claim their method surpasses existing automated segmentation algorithms. Focal brain lesions are a critical diagnostic marker for MS and also serve as valuable indicators for evaluating treatment response and predicting future disability. The corresponding author, Francesco La Rosa, PhD, from the Icahn School of Medicine at Mount Sinai Health System, and colleagues emphasize that the number and volume of lesions are essential data points for MS research.
Reliable lesion segmentation lays the groundwork for exploring advanced imaging biomarkers, such as central vein sign and paramagnetic rim lesions, the team notes.
Traditionally, MS lesions can be identified on T2-weighted MRI scans. However, manual segmentation is time-consuming and prone to variability between different raters. Advances in machine learning and AI have enabled automated segmentation, with numerous approaches developed and even an international competition last year in which Mount Sinai participated.
La Rosa and colleagues argue that the most successful automated methods are built on convolutional neural networks. Yet they point out that the majority of automated systems still rely on both T2-weighted FLAIR (FLAIR) images and T1-weighted scans. Many new algorithms are trained on high-quality research scans that aren’t always available in everyday clinical practice.
To address this gap, the team aimed to create an automated method effective with only T2w FLAIR images. They built a deep-learning model using 668 MS patient scans, acquired on 1.5T and 3T MRI machines, and named it FLAMeS — FLAIR Lesion Analysis in Multiple Sclerosis.
In their evaluation, the researchers compared FLAMeS to two publicly available benchmark algorithms using both research-grade and clinical-grade MRI data. The benchmarks were the Lesion Segmentation Toolbox (LST) methods and the Sequence Adaptive Multimodal SEGmentation (SAMSEG) algorithm. LST itself comprises three distinct methods; the team used one when only a FLAIR image was available and a second, AI-based method when both FLAIR and T1-weighted scans were present.
To test performance, the study used three external datasets. Evaluation included qualitative assessments by two blinded experts and quantitative comparisons between automated segmentations and ground-truth lesion masks using standard metrics.
In qualitative reviews of 20 scans, one expert judged FLAMeS as the most accurate in 15 cases, and the other in 17 cases. Quantitatively, FLAMeS outperformed the other methods, achieving a higher positive predictive rate and a lower false-positive rate. The authors note that the competing methods missed both large and small lesions, whereas FLAMeS only missed lesions smaller than about 10 mm3.
The researchers acknowledge that FLAMeS’ superior performance could stem from its advanced underlying architecture, stronger training data, or a combination of both. They express optimism that the improved accuracy and robustness to image quality variations, resolution differences, and acquisition protocols make FLAMeS a valuable asset for MS research.
Importantly, the team has publicly released the model to encourage others to fine-tune and apply it in different settings, potentially accelerating MS research and clinical workflows.
Key points to remember:
- FLAMeS is designed to work with only T2w FLAIR images, avoiding the need for T1-weighted scans.
- The method demonstrated superior accuracy in both qualitative and quantitative assessments compared with established benchmarks.
- The potential benefits include faster lesion analysis, better consistency across observers, and improved integration with emerging imaging biomarkers.
Follow-up questions for readers:
- Do you think FLAMeS could meaningfully change clinical practice, or will it primarily affect research workflows in its early stages?
- What are the implications of relying on a single imaging modality for lesion segmentation in diverse clinical settings?
- How might this approach influence the development or validation of new MS biomarkers in the near term?
