Quantificação da tortuosidade de vasos sanguíneos: definições e influência das etapas de realce e segmentação

Abstract

Morphological features of blood vessels can be used as markers of several pathologies affecting the tissues of organisms. Among these features, tortuosity displays a vital role in the study and diagnostic of conditions that affect blood flow. Advances in computational power and imaging techniques have made it possible to develop automatic methodologies for quantifying tortuosity in images with hundreds or even thousands of blood vessels. These methodologies usually involve the enhancement, segmentation, and skeletonization of blood vessels, each of which has been studied in several works in the literature. Frequently, these works ignore the influence that blood vessel tortuosity has on the quality of the result. In particular, many of the developed algorithms are not focused on processing blood vessels with high tortuosity, which tend to be relevant for several diseases. As a result, systematic problems (or biases) on quantifying blood vessel tortuosity can occur, leading to wrong conclusions regarding the studied vessels. Thus, we studied the influence that tortuosity has on different stages of digital processing of blood vessel images, especially regarding the steps of enhancement and segmentation. Experiments performed with an artificial blood vessel generator suggest that biases may occur when using traditional enhancement algorithms in vessels with high tortuosity. Using real images of blood vessels in the retina, we identified that databases mainly composed of tortuous vessels may affect the quality of segmentation by deep learning methods, with possible overestimation of the average tortuosity value of a vessel network. The experiments with deep learning were expanded with a large dataset of confocal microscopy images of the mouse cortex. We found possible biases that can occur when using pre-trained off-the-shelf networks to perform morphometric analyses of blood vessels. In addition, we identified the possibility of mitigating part of those biases by a fine-tuning step performed with new datasets, or by expanding the training dataset with transformations applied to the original data. It is expected that the proposed analyzes will lead to an improvement in the accuracy of morphometric values obtained in frontier research regarding the cardiovascular system.