CardioSmartAssist: A customisable AI framework for echocardiography-based cardiac assessment

Background and Objective. Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, accounting for approximately 20.5 million deaths annually, nearly one-third of all global deaths. Despite its central role in cardiology, echocardiographic assessment remains subject to significant inter- and intra-observer variability, particularly based on manual frame selection and segmentation. This limitation has driven increasing interest in Deep Learning (DL) solutions capable of enabling more objective, reproducible, and efficient analyses. Methods. This paper introduces CardioSmartAssist, a deep learning framework for automatic left ventricular segmentation and multi-cycle ejection fraction estimation relying solely on echocardiographic videos. The system integrates frame-by-frame segmentation visualisation, anomaly detection, and volume tracking to enhance clinical usability. Moreover, a key feature is its continuous learning mechanism, which allows clinician-corrected segmentations to be stored and used for progressive model refinement. Results. The framework is based on a MultiResUNet architecture, trained on public (EchoNet-Dynamic) and proprietary (CardioSmartSet) datasets, achieving Dice Coefficient Scores of 0.9328 and 0.9189, respectively. On the held-out test set, the EF estimated by the system showed a mean absolute difference of 10\% compared with clinically reported EF values, which is lower than the typical inter-operator variability of approximately 13%. Conclusion. CardioSmartAssist resulted to be a promising tool for consistent cardiac evaluations, improving access to diagnostics, and enhancing clinical decision-making through smart assistance.