MRI‐Based Manual versus Automated Corpus Callosum Volumetric Measurements in Multiple Sclerosis

Abstract

BACKGROUND AND PURPOSE Corpus callosum atrophy is a neurodegenerative biomarker in multiple sclerosis (MS). Manual delineations are gold standard but subjective and labor intensive. Novel automated methods are promising but require validation. We aimed to compare the robustness of manual versus automatic corpus callosum segmentations based on FreeSurfer. METHODS Nine MS patients (6 females, age 38 ± 13 years, disease duration 7.3 ± 5.2 years) were scanned twice with repositioning using 3‐dimensional T1‐weighted magnetic resonance imaging on three scanners (two 1.5 T and one 3.0 T), that is, six scans/patient, on the same day. Normalized corpus callosum areas were measured independently by a junior doctor and neuroradiologist. The cross‐sectional and longitudinal streams of FreeSurfer were used to segment the corpus callosum volume. RESULTS Manual measurements had high intrarater (junior doctor .96 and neuroradiologist .96) and interrater agreement (.94), by intraclass correlation coefficient (P < .001). The coefficient of variation was lowest for longitudinal FreeSurfer (.96% within scanners; 2.0% between scanners) compared to cross‐sectional FreeSurfer (3.7%, P = .001; 3.8%, P = .058) and the neuroradiologist (2.3%, P = .005; 2.4%, P = .33). Longitudinal FreeSurfer was also more accurate than cross‐sectional (Dice scores 83.9 ± 7.5% vs. 78.9 ± 8.4%, P < .01 relative to manual segmentations). The corpus callosum measures correlated with physical disability (longitudinal FreeSurfer r = –.36, P < .01; neuroradiologist r = –.32, P < .01) and cognitive disability (longitudinal FreeSurfer r = .68, P < .001; neuroradiologist r = .64, P < .001). CONCLUSIONS FreeSurfer's longitudinal stream provides corpus callosum measures with better repeatability than current manual methods and with similar clinical correlations. However, due to some limitations in accuracy, caution is warranted when using FreeSurfer with clinical data.

Comparison of Manual Cross-Sectional Measurements and Automatic Volumetry of the Corpus Callosum, and Their Clinical Impact: A Study on Type 1 Diabetes and Healthy Controls

“…The variable error thus introduced causes the correlation of CCA vs. CCV to suffer. FreeSurfer version 6.0 was however found to correlate well with manual measurement using the longitudinal stream by Platten et al (9). The correlation was worse using the cross-sectional stream as is the case in the present study.…”

“…Repeated measurements in the same slice using standard PACS area-tracing tools gave highly uniform results suggesting there is no advantage to multiple averaged measurements over single measurements. CCA measurements have been found to be highly repeatable regardless of operator experience (9).…”

“…The correlation between CCA and CCV or MISS and eTIV has been studied in a smaller material. The agreement between measurements made by junior physicians and experienced neuro-radiologists was found to be excellent, which means that CCA can be easily measured in routine clinical practice (9). Not all imaging centers are equipped to routinely perform volumetry, and simple manual measurements could thus prove valuable.…”

Deep Learning Corpus Callosum Segmentation as a Neurodegenerative Marker in Multiple Sclerosis

“…Previous studies employing corpus callosum segmentations have reported varying levels of success. Manual segmentations have been shown useful and are often used as gold standard when comparing other methods 9,12 . However, manual measurements are tedious and have an inherent inter‐ and intrarater variability.…”

“…There are several methods for segmenting the corpus callosum based on MRI, including manual segmentations, 9 atlas‐based software, such as FreeSurfer, 10 and machine learning algorithms 11 . FreeSurfer can render 3D segmentations of the corpus callosum with high precision but lower accuracy, as compared to manual delineations 12 . In recent years, artificial intelligence has become more widely applied in radiology and convolutional neural networks, a subclass of deep learning inspired by human layered neuronal connectivity, are particularly useful in segmenting radiological images 13 .…”