Contributions
Abstract: P1133
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 25 Biomarkers
Objective: Lesion volume and atrophy changes are important quantitative imaging biomarkers of respective inflammatory and neurodegenerative processes, characteristic for patients with multiple sclerosis (MS). These biomarkers may be extracted from brain magnetic resonance (MR) images using various analysis methods, which generally yield different biomarker values, also depending on MR image quality. Using a novel reference-free regression framework we address the questions of how accurate and precise is each of the measurement methods, without the need of reference measurements.
Subjects and methods: The MR images of 48 patients diagnosed with relapsing remitting MS (40.8±12.3 years old, 34 females) were acquired on two 3T scanners using a harmonized imaging protocol. For 26 patients a follow-up MR scans were acquired 12 weeks after the baseline scan. On baseline scans lesion masks were obtained by three untrained and one trained automated methods, while reference masks used for validation were created by three neuroradiologists using semi-automated approach. For atrophy changes four different methods were applied on the follow-up scans, namely the Siena method, brain parenchymal volume fraction (BPVC), brain shift integral (BSI) and deformation Jacobian integration (JI). Computed lesion volumes and atrophy were then fed into the reference-free regression framework that characterized systematic and random errors of each method.
Results: Rankings of lesion segmentation methods with respect to root mean square error (RMSE) compared to reference volumes and standard deviation (STD) of random volume errors were consistent with the results of least squares regression against the reference. For atrophy changes the reference values were not available, thus RMSE was computed against true values estimated by the framework. Regression framework identified the trained lesion segmentation as most accurate and precise (RMSE=6.5 ml, STD=1.1 ml). Most accurate atrophy change method was BSI (RMSE=0.24%), while JI was found most precise (STD=0.09%).
Conclusion: Accuracy and precision of lesion segmentation methods as computed by the reference-free regression was consistent with gold standard based assessment, while corresponding estimates for atrophy changes were in good agreement with previous research reports. The framework is a promising tool that may be widely applicable for identification of optimal biomarker extraction methods in various domains.
Disclosure: Hennadii Madan: nothing to disclose. Lina Savsek: nothing to disclose. Sasa Sega Jazbec: nothing to disclose. Ziga Spiclin: nothing to disclose.
Abstract: P1133
Type: Poster
Abstract Category: Pathology and pathogenesis of MS - 25 Biomarkers
Objective: Lesion volume and atrophy changes are important quantitative imaging biomarkers of respective inflammatory and neurodegenerative processes, characteristic for patients with multiple sclerosis (MS). These biomarkers may be extracted from brain magnetic resonance (MR) images using various analysis methods, which generally yield different biomarker values, also depending on MR image quality. Using a novel reference-free regression framework we address the questions of how accurate and precise is each of the measurement methods, without the need of reference measurements.
Subjects and methods: The MR images of 48 patients diagnosed with relapsing remitting MS (40.8±12.3 years old, 34 females) were acquired on two 3T scanners using a harmonized imaging protocol. For 26 patients a follow-up MR scans were acquired 12 weeks after the baseline scan. On baseline scans lesion masks were obtained by three untrained and one trained automated methods, while reference masks used for validation were created by three neuroradiologists using semi-automated approach. For atrophy changes four different methods were applied on the follow-up scans, namely the Siena method, brain parenchymal volume fraction (BPVC), brain shift integral (BSI) and deformation Jacobian integration (JI). Computed lesion volumes and atrophy were then fed into the reference-free regression framework that characterized systematic and random errors of each method.
Results: Rankings of lesion segmentation methods with respect to root mean square error (RMSE) compared to reference volumes and standard deviation (STD) of random volume errors were consistent with the results of least squares regression against the reference. For atrophy changes the reference values were not available, thus RMSE was computed against true values estimated by the framework. Regression framework identified the trained lesion segmentation as most accurate and precise (RMSE=6.5 ml, STD=1.1 ml). Most accurate atrophy change method was BSI (RMSE=0.24%), while JI was found most precise (STD=0.09%).
Conclusion: Accuracy and precision of lesion segmentation methods as computed by the reference-free regression was consistent with gold standard based assessment, while corresponding estimates for atrophy changes were in good agreement with previous research reports. The framework is a promising tool that may be widely applicable for identification of optimal biomarker extraction methods in various domains.
Disclosure: Hennadii Madan: nothing to disclose. Lina Savsek: nothing to disclose. Sasa Sega Jazbec: nothing to disclose. Ziga Spiclin: nothing to disclose.