Abstract
Primary myelofibrosis (PM) is a chronic blood cancer which increases burden on the spleen to produce blood cells and results in palpable splenomegaly. In the clinic, splenomegaly is classified based on the distance between the spleen's lowest point and the left costal margin, however, this method is highly subjective and depends on the subject's position and respiration. Imaging techniques have the potential to provide accurate, reliable, and reproducible measurements of splenic volume (SV). In clinical trials assessing therapy response, an accepted imaging-based endpoint is ≥35% reduction in SV at week 24 from baseline as measured by Magnetic Resonance Imaging (MRI) or Computer Tomography (CT). A ≥25% increase in SV is typically considered progression. The most accurate method for volume assessment is manual segmentation, since the entire spleen boundary can be utilized for the volume calculation. This study compared two other volume estimation methods: ellipsoid method and a model proposed by Bezerra et al (AJR Am J Roentgenol. 2005). We compared the methods' performance in assessing treatment response or progression based on SV change from baseline to week 24.
Imaging data from 30 participants were used in this study, predominantly acquired using MRI modality; CT was used as an alternative, when MRI was contraindicated. Scans from two timepoints per participant were used: baseline and 24 weeks after start of treatment. For the manual segmentation method, preliminary regions of interest were manually outlined on every imaging slice by an experienced imaging analyst and then reviewed by a trained radiologist. SV was derived by multiplying the number of voxels contained in the spleen outlined by the voxel size of the scan. For the ellipsoid method, maximum width (W) and orthogonal thickness (T) were measured on the axial images. Length (L) was measured by multiplying the number of slices containing spleen by the slice interval. Ellipsoid volume was calculated as follows:
V = W * T * L * π / 6
For the length-estimated SV based on the Bezerra et al model, spleen length was utilized as shown:
V = (L - 5.8006) / 0.0126
For each of the three methods, percent change in SV was calculated from baseline to week 24. Pearson's correlation coefficient and Bland Altman analysis were implemented for comparison of methods to manual segmentation. Sensitivity and specificity analysis was performed to determine the accuracy of each method to predict response or progression.
The manual segmentation volume was significantly correlated with both the ellipsoid method (r(58) = 0.94, p < 0.0001) and the length-estimated method (r(58) = 0.89, p < 0.0001). When assessing percent changes from baseline to week 24 using manual segmentation, 4 of the participants achieved splenic response and 4 progressed with 25% increase in SV. However, analysis using ellipsoid method yielded 3 responding and 2 progressing participants. Finally, analysis with length-estimated volume yielded no responding or progressing participants. This data is also illustrated in Table 1, which shows the sensitivity and specificity results. Figure 1 illustrates Bland-Altman plots, suggesting that ellipsoid method provides a more accurate estimation of the change in SV compared to length-estimated volume. Furthermore, we found that the inaccuracy with length-estimated volume increases with larger spleens (not shown).
Change in spleen volume contributes to the primary/ secondary endpoints in large multi-center clinical trials for myelofibrosis, so it is imperative that the methods used to measure SV are consistent across imaging sites. The current standard for assessing SV is the manual segmentation method because it provides the most comprehensive measurement of spleen size however, this process is burdensome, time consuming, and requires specific training. The ellipsoid and length-estimated methods were strongly correlated with the manual segmentation method; however, they were not as sensitive when determining treatment response or progression. The length-estimated method had the least level of agreement with manual segmentation. The ellipsoid method may be a better alternative; however, it is important to use one method consistently across all visits for a study participant. Additional work is required to test performance of methods on a larger cohort, as well as assess automated segmentation algorithms that may reduce the burden of manual tracing.
Figure 1 Figure 1.
Disclosures
No relevant conflicts of interest to declare.
Ellipsoid method for convex stochastic optimization in small dimension
