Automatic measurement of temporalis muscle thickness from CT head scans using deep learning

Abstract

ObjectiveTo construct and validate a deep-learning (DL) model for the automatic quantification of temporalis muscle thickness (TMT) in CT head scans. Materials and methodsWe developed and evaluated the performance of a DL-based method for the measurement of temporalis muscle thickness using publicly available CT head scans. Reference standard TMT was established using a previously published measurement protocol, applied on 198 CT head scans obtained from a publicly available database, originally collected from various radiology centers within one city in 2017. A DL landmark detection model was trained to measure the temporalis muscle thickness. The absolute error and correlation between DL-based and reference standard TMT measurements were calculated. Additionally, the ability of the DL-based measurements to stratify subjects into low TMT vs normal TMT groups was assessed using the metrics of specificity, sensitivity, and accuracy. ResultsThe median reference TMT value was 6.0 mm (95% CI: 5.4, 6.5); the median DL-based TMT value was 5.8 mm (95% CI: 5.6, 6.3). The mean absolute error for TMT was 0.7 mm (95% CI: 0.6, 0.9). The correlation coefficient between reference and DL-based TMT was 0.9 (95% CI: 0.8, 0.9). The DL-based measurements classified the patients into low and normal TMT groups with sensitivity of 84.2%, specificity of 85.0% and accuracy of 84.6%. ConclusionsOur DL-based pipeline allows for fully automated and reproducible quantification of temporal muscle thickness and patient stratification into low and normal TMT groups. Summary statementIn this study we developed and tested a deep learning model for the measurement of temporal muscle thickness in CT head scans, showing potential utility for clinical decision making. Key pointsO_LITemporalis muscle thickness surrogates muscle mass, which is associated with frailty; however, there is no automatic tool to measure temporalis muscle thickness in CT head scans C_LIO_LIOur deep learning model-based temporalis muscle thickness measurements presented a strong, positive linear correlation with reference standard measurements, and less than 1mm mean absolute error C_LIO_LIThe deep learning-based measurements demonstrated strong potential for stratifying subjects into low versus normal muscle levels, enabling quick assessment of temporalis muscle status. C_LI