Abstract Background While coronary artery tortuosity (CAT) has been observed in clinical practice, the evidence supporting its association with vascular disease and complications remains limited. Currently, CAT assessment primarily relies on qualitative visual inspection, leading to inter- and intra-operator variability. Moreover, the absence of standardized, clinically validated classification criteria makes the objective quantification of CAT challenging. Purpose This study introduces a novel unsupervised approach based on Hidden Markov Models that exploits geometric features of the vessel’s centerline to identify tortuosity patterns along the course of coronary arteries. Methods Coronary Computed Tomography Angiography (CCTA) images were used to reconstruct a three-dimensional model of the vessel and extract the centerline. The employed dataset included CCTA scans from 319 patients undergoing imaging for clinical purposes. The study focused on the left anterior descending artery (LAD), analyzing end-diastolic acquisitions to ensure consistency. A sliding window was moved along the centerline to compute geometric features such as curvature and local tortuosity, providing a comprehensive representation of vessel morphology. These features were encoded and used as input for the model. A multivariate discrete Hidden Markov Model (HMM) was developed to classify each point along the centerline into one of three tortuosity levels: absent/low, moderate, or severe. Once trained, the HMM assigns each point to the most likely tortuosity level based on the encoded features. The model's performance was assessed by comparing its predictions with ground truth annotations provided by an independent expert who manually classified 18 LAD arteries into linear, moderate and tortuous regions using the 3D models. To enhance the explainability of the method, multiple HMMs were trained using all the combinations of features. By analyzing the outcomes of these models, the optimal feature set was identified, excluding those that had minimal impact on the performance. Results The best HMM demonstrated good agreement with expert annotations (precision=0.87, recall=0.86, accuracy: low=0.93, moderate=0.83, severe=0.74), providing a reproducible method for CAT quantification. The model enables the automated spatial localization of varying tortuosity levels along the artery, facilitating objective assessment and highlighting regions of potential clinical significance. Conclusions This method offers an explainable and data-driven alternative to qualitative evaluations. However, the absence of standardized thresholds for CAT classification complicates validation, as expert annotations inherently involve a degree of subjectivity. Future work should focus on validation studies to establish the prognostic value of this automated tool, potentially enabling its integration into routine clinical decision-making pathways for improved cardiovascular risk stratification.
Ferrari et al. (Sat,) studied this question.