Validation of a semi-automated method to quantify lesion volume changes in multiple sclerosis on 2D proton-density-weighted scans based on image subtraction

Background: The detection and quantification of changes in white matter lesions in the brain is important to monitor treatment effects in patients with multiple sclerosis (MS). Existing automatic tools predominantly require FLAIR images as input which are not always available, or only focus on new/enlarging activity. Therefore, we developed and validated a semi-automated method to quantify lesion volume changes based on 2D proton-density (PD)-weighted images and image subtraction. This semi-automated method provides insight in both “positive” activity (defined as new and enlarging lesions) and “negative” activity (disappearing and shrinking lesions). Methods: Yearly MRI scans of patients with early MS from the REFLEX/REFLEXION studies were used. The maximum follow-up period was 5 years. Two PD-weighted images were normalized, registered to a common halfway-space, intensity-matched, and subsequently subtracted. Within manual lesion masks, lesion changes were quantified using a subtraction intensity threshold and total lesion volume change (TLVC) was calculated. Reproducibility was measured by assessing transitivity, specifically, we calculated the intraclass correlation coefficient for the absolute agreement (ICCtrans) and the difference (Δtrans) between the direct one-step and indirect multi-step measurements of TLVC between two visits. Accuracy was assessed by calculating both the intraclass correlation coefficient for absolute agreement (ICCacc) and the difference (Δacc) between the one-step semi-automated TLVC and manually measured lesion volume change (numerical difference) between two visits. Spearman's correlations (rs) were used to assess the relation of global and central atrophy, manually measured T2 lesion volume, and lesion volume change with the method's performance as reflected by the difference measures |Δtrans| and Δacc. An alpha of 0.05 was used as the cut-off for significance. Results: Reproducibility was excellent, with ICCtrans values ranging from 0.90 to 0.96. Accuracy was good overall, with ICCacc values ranging from 0.67 to 0.86. The standard deviation of Δtrans ranged from 0.25 to 0.86 mL. The mean of Δacc ranged from 0.11 to 0.37 mL and was significantly different from zero. Both global and central atrophy significantly correlated with lower reproducibility (correlation of |Δtrans| with global atrophy, rs = −0.19 to −0.28, and correlation of |Δtrans| with central atrophy, rs = 0.22 to 0.34). There was generally no significant correlation between global/central atrophy and accuracy. Higher lesion volume was significantly correlated with lower reproducibility (rs = 0.62). Higher lesion volume change was significantly correlated with lower reproducibility (rs = 0.22) and lower accuracy (correlation of Δacc with lesion volume change, rs = −0.52). Discussion: The semi-automated method to quantify lesion volume changes has excellent reproducibility and overall good accuracy. The amount of atrophy and especially lesion volume (change) should be taken into account when applying this method, as an increase in these variables might affect the quality of the results. Conclusion: Overall, the semi-automated subtraction method allows a valid and reliable quantitative investigation of lesion volume changes over time in (early) MS for follow-up periods up to 5 years.