Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RimSet: Quantitatively Identifying and Characterizing Chronic Active Multiple Sclerosis Lesion on Quantitative Susceptibility Maps

Jinwei Zhang, Thanh D. Nguyen, Renjiu Hu, Susan A. Gauthier, Yi Wang, Hang Zhang

TL;DR

RimSet tackles the challenge of quantitatively identifying and characterizing Rim+ chronic active MS lesions on Quantitative Susceptibility Mapping (QSM). It combines RimSeg, a regularized level-set–based rim segmentation method, with an extensive radiomic and Local Binary Pattern texture feature set, and trains an XGBoost classifier to distinguish rim+ from rim- lesions and to quantify rim properties. On simulated data and a large in vivo MS dataset, RimSet achieves state-of-the-art performance (e.g., ROC AUC up to $0.971$, PR AUC $0.737$, and subject-level Pearson ρ up to $0.91$ with MSE $0.85$), outperforming RimNet, APRL, and QSMRimNet. The approach provides interpretable, quantitative biomarkers with potential to standardize rim+ lesion assessment across centers and support clinical decision-making, while remaining adaptable and real-time capable through RimSeg and a compact feature set. Limitations include challenges with partial rims and differentiating rim+ from rim- lesions in complex myelin/iron contexts, motivating future work with additional transforms and consensus validation on real data.

Abstract

Background: Rim+ lesions in multiple sclerosis (MS), detectable via Quantitative Susceptibility Mapping (QSM), correlate with increased disability. Existing literature lacks quantitative analysis of these lesions. We introduce RimSet for quantitative identification and characterization of rim+ lesions on QSM. Methods: RimSet combines RimSeg, an unsupervised segmentation method using level-set methodology, and radiomic measurements with Local Binary Pattern texture descriptors. We validated RimSet using simulated QSM images and an in vivo dataset of 172 MS subjects with 177 rim+ and 3986 rim-lesions. Results: RimSeg achieved a 78.7% Dice score against the ground truth, with challenges in partial rim lesions. RimSet detected rim+ lesions with a partial ROC AUC of 0.808 and PR AUC of 0.737, surpassing existing methods. QSMRim-Net showed the lowest mean square error (0.85) and high correlation (0.91; 95% CI: 0.88, 0.93) with expert annotations at the subject level.

RimSet: Quantitatively Identifying and Characterizing Chronic Active Multiple Sclerosis Lesion on Quantitative Susceptibility Maps

TL;DR

RimSet tackles the challenge of quantitatively identifying and characterizing Rim+ chronic active MS lesions on Quantitative Susceptibility Mapping (QSM). It combines RimSeg, a regularized level-set–based rim segmentation method, with an extensive radiomic and Local Binary Pattern texture feature set, and trains an XGBoost classifier to distinguish rim+ from rim- lesions and to quantify rim properties. On simulated data and a large in vivo MS dataset, RimSet achieves state-of-the-art performance (e.g., ROC AUC up to , PR AUC , and subject-level Pearson ρ up to with MSE ), outperforming RimNet, APRL, and QSMRimNet. The approach provides interpretable, quantitative biomarkers with potential to standardize rim+ lesion assessment across centers and support clinical decision-making, while remaining adaptable and real-time capable through RimSeg and a compact feature set. Limitations include challenges with partial rims and differentiating rim+ from rim- lesions in complex myelin/iron contexts, motivating future work with additional transforms and consensus validation on real data.

Abstract

Background: Rim+ lesions in multiple sclerosis (MS), detectable via Quantitative Susceptibility Mapping (QSM), correlate with increased disability. Existing literature lacks quantitative analysis of these lesions. We introduce RimSet for quantitative identification and characterization of rim+ lesions on QSM. Methods: RimSet combines RimSeg, an unsupervised segmentation method using level-set methodology, and radiomic measurements with Local Binary Pattern texture descriptors. We validated RimSet using simulated QSM images and an in vivo dataset of 172 MS subjects with 177 rim+ and 3986 rim-lesions. Results: RimSeg achieved a 78.7% Dice score against the ground truth, with challenges in partial rim lesions. RimSet detected rim+ lesions with a partial ROC AUC of 0.808 and PR AUC of 0.737, surpassing existing methods. QSMRim-Net showed the lowest mean square error (0.85) and high correlation (0.91; 95% CI: 0.88, 0.93) with expert annotations at the subject level.
Paper Structure (21 sections, 3 equations, 7 figures, 4 tables)

This paper contains 21 sections, 3 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Datasets
  4. Numerical Simulation Data
  5. In Vivo MS Data
  6. Segmentation and Measurements
  7. Active Contour Model
  8. Regularized Level-Set for Rim Segmentation
  9. Quantitative Measurements
  10. Comparator Methods
  11. Implementation Details
  12. Statistical Analysis
  13. Lesion-wise Analysis
  14. Subject-wise Analysis
  15. Measurement Analysis
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Visual representation of the diversity of simulated rim+ and rim- lesions, including their responses to varying noise levels and complex structural characteristics. The first row (a) illustrates variations in the radius, rim thickness, presence of partial rims, oval shape, and vein intersections in the rim+ lesions. The second row (b) displays the corresponding rim segmentation masks derived from Eq. \ref{['eq:levelset-new']}, demonstrating its effectiveness in distinguishing rims even in challenging scenarios such as vein intersections. The third row (c) represents rim+ lesions as full shells with increasing noise levels (columns 1-4) and rim- lesions as solid spheres with similar noise augmentation (columns 5-8).
  • Figure 2: A visual illustration of volume fraction of simulated rim+ lesions. The density histogram exhibits a right-skewed, or long-tailed distribution, indicating a diversity of smaller volume fractions and a relatively fewer number of larger fractions.
  • Figure 3: A measurement importance plot for the identification of rim+ lesions, derived from Xgboost. A total of 24 measurement, spanning a variety of intensity, texture, shape, and statistical properties, were evaluated. The 'Mean-Distance' was found to be the most influential measurement with a score of 905, followed by 'LBP' and 'Kurtosis' with scores of 847 and 742 respectively. On the other hand, 'Mean' showed the least influence with a score of 107.8. The measurement importance scores offer insights into the discriminative capabilities of these features in the characterization of rim+ lesions.
  • Figure 4: Examples of misclassified lesions in the simulated dataset.
  • Figure 5: Performance evaluation of RimSeg on simulated data. (a) presents the robustness of RimSeg against noise by displaying the Dice score trends for rim+ lesions with varying levels of Gaussian noise (sigma). The comparison between full rim and partial rim lesions is also highlighted, showing a performance drop when partial rims are present. (b) provides examples of simulated rim+ and rim- lesions with increasing levels of Gaussian noise ($\sigma=1.0, 3.0, 5.0, 7.0$). The effects of increased noise levels on the lesion segmentation by the proposed RimSeg method are illustrated, with over-segmentation observed at a $\sigma=7.0$.
  • ...and 2 more figures