Machine learning and fractal-based analysis for the automated diagnosis of cardiovascular diseases using magnetic resonance

Abstract

[en] Cardiac magnetic resonance (CMR) is the reference imaging modality for the diagnose of cardiovascular diseases. Traditionally, simple CMR parameters related to the volume and shape of the cardiac structures are calculated by the medical professionals by means of manual or semi-automated approaches. This process is time-consuming and prone to human errors. Moreover, despite the importance of these traditional CMR indexes, they often fail to fully capture the complexity of the cardiac tissue. In this work, we propose a novel approach for automated cardiovascular disease diagnosis, using ischemic heart disease as an example use case. Towards this aim, we will use a state-of-the-art technology, supervised machine learning, and a promising mathematical tool, fractal-based analysis. In order to undertand the potential information that can be derived from fractal-based features, we introduce and explore the concepts of Haussdorff dimension, box-counting dimension and lacunarity. We describe the interrelationships among these concepts and present computational algorithms for calculating box-counting dimension and lacunarity. The study is based on data from a large-cohort study, UK Biobank, to extract box-counting dimension and lacunarity from CMR textures focusing on three cardiac structures of medical interest: the left ventricle, the right ventricle and the myocardium. The extraction of these features allows us to obtain quantitative parameters regarding the complexity and heterogeneity of the tissue. These fractal features, both individually and in conjunction with other vascular risk factors and CMR traditional indexes, are employed as inputs to state-of-the-art machine learning models, including SVM, XGBoost, and random forests. The objective is to determine if the inclusion of fractal features enhances the performance of currently employed parameters. The performance evaluation of our models is based on metrics such as balanced accuracy, F1 score, precision, and recall. The results obtained demonstrate the potential of fractal-based features in improving the accuracy and reliability of cardiovascular diseases diagnosis.