Automated CIMT Measurement from Ultrasound Using Deep Learning with Uncertainty Estimation

Carotid intima-media thickness (CIMT) is a widely used marker for cardiovascular risk assessment, but manual measurement from ultrasound images is time-consuming and subject to substantial inter-observer variability. We propose LUCID - a single-stage deep learning pipeline combining a U-Net with a pretrained ResNet34 encoder for segmentation, sub-pixel boundary extraction for CIMT computation, and Monte Carlo Dropout with post-hoc calibration for uncertainty estimation. Trained on only 500 expert-annotated images from the Carotid Ultrasound Boundary Study (CUBS) benchmark using five-fold cross-validation, the model achieves 0.142 mm mean absolute error, matching the best traditional method by Consiglio Nazionale delle Ricerche (CNRIT, 0.139 mm) while requiring no task-specific preprocessing. The calibrated uncertainty estimation feeds a triage system that automatically accepts confident predictions and flags uncertain cases for clinical review. This is the first CIMT measurement method to integrate calibrated uncertainty estimation, enabling safer deployment in clinical screening workflows.