Abstract WP246: Automated sarcopenia assessment using routinely acquired CT angiography in stroke and transient ischemic attack

Background: Sarcopenia, defined by reduced skeletal muscle mass and function, is highly prevalent in older adults and contributes to disability, frailty, and adverse outcomes across chronic and acute diseases. In stroke, pre-existing sarcopenia is linked with poorer recovery, yet it is rarely assessed in practice. Neck CT angiography (CTA), routinely acquired for suspected stroke and transient ischemic attack (TIA), captures the third cervical vertebral level (C3), where skeletal muscle index (SMI) can be estimated. We aimed to develop and validate a fully automated pipeline for sarcopenia assessment using CTA. Methods: We analyzed 260 consecutive CTA head and neck scans from a population-based cohort of patients presenting with stroke/TIA in Alberta, Canada. At the C3 level, right and left sternocleidomastoid and paravertebral muscles were manually segmented by radiologist-supervised trained assessors using ITK-SNAP within the skeletal muscle density range (−29 to 150 HU) to establish ground truth. Cases were divided 80/20 into training and testing datasets. The fully automated pipeline consisted of C3 slice detection using a decision-making algorithm grounded in real physical coordinates, followed by a novel segmentation framework based on state-of-the-art diffusion models for target muscle segmentation (Flowchart in Figure 1 ). The volumes of the segmented muscles are automatically calculated, allowing SMI to be determined using a standard height/weighted-standardized equation. We evaluated accuracy of C3 localization using standard detection performance metrics and assessed muscle segmentation for sarcopenia evaluation with the Dice coefficient. Results: 208 cases were used for training and 52 for testing (Exemplars in Figure 2 ). In testing data, the fully automated algorithm achieved excellent C3 slice detection performance with a correct recognition rate of 0.846 (95% CI:0.719-0.931), allowing 10mm vertical error tolerance, with a mean absolute error of 1.99mm. Muscle segmentation was highly accurate (Overall Dice 0.91 (95%CI:0.904-0.916)). Conclusions: Our automated model for muscle segmentation for sarcopenia evaluation demonstrated robust performance in this routine stroke care cohort. This will enable opportunistic evaluation of any patient undergoing CTA. Future work will apply this automated evaluation for stroke outcome prediction in large datasets.