Modified UNet-enhanced ultrasonic superb microvascular imaging feature extraction and grading of carpal tunnel syndrome

Carpal tunnel syndrome (CTS) is recognized as the most frequently encountered median nerve (MN) entrapment neuropathy, with a disproportionate burden in middle-aged and elderly individuals and in occupational groups with repetitive wrist use. Anatomically, CTS is characterized by compression of the median nerve within the confined space between the transverse carpal ligament and flexor tendons, and microcirculatory impairment is regarded as one of its key pathological bases. Although electrodiagnostic assessments are considered as diagnostic gold standard, their utility is limited by suboptimal patient compliance, procedural discomfort, and inadequate sensitivity for detecting mild disease. This study integrates ultrafast Superb Microvascular Imaging (SMI) with a classification-guided, improved UNet segmentation modal and quantitative image analysis to objectively extract microvascular features for CTS grading. In a cohort of 105 patients (21 mild, 71 moderate, 13 severe CTS) and 21 healthy controls, longitudinal and transverse SMI cine loops were segmented using an improved UNet with cross-plane classification guidance. The modified network can yielded superior segmentation effect over a traditional UNet. From segmented regions we extracted 6 SMI-derived geometric features, which were then used as predictors in a nonlinear quadratic regression model for CTS severity grading. The model achieved 93.7 % overall classification accuracy and an AUC of 0.95 in cross validation. Independent blind validation (n = 12) showed strong agreement with expert sonographers (Kappa = 0.87). These results demonstrate that high spatiotemporal SMI combined with anatomy-aware deep learning model could enable reproducible extraction of microvascular geometry, and supports robust, noninvasive grading of CTS, with potential for deployment on portable ultrasound platforms for point-of-care screening and bedside ultrasonic monitoring.