Left ventricle segmentation in Cardiac Magnetic Resonance Imaging is important to obtain biomarkers for heart disease diagnosis. In clinical routine, these exams, comprising a large amount of images, are usually manually or semi-manually segmented by experts, demanding repetitive effort and impacting the quality of diagnosis. Several approaches for automatic segmentation have been proposed, mainly using methods based on deep learning and energy/cost minimization, such as graphs, deformable models and atlas. On average, deep learning methods have obtained results close to those of experts, but still produce many anatomical errors. In order to mitigate this limitation, recent hybrid approaches have combined deep learning and minimization methods, offering greater precision. However, the way these approaches are combined is still limited, since there is no automatic adaptation of the strategies that aim to reduce the specific errors made by each one. Considering this problem, this PhD project aims to build an adaptive hybrid approach that combines deep learning and energy/cost minimization methods for automatic segmentation of the left ventricle in Cardiac Magnetic Resonance Imaging. Automatic adaptation favors the methods to work together and can improve the result in unusual cases. The results will be evaluated by comparing the automatic segmentations with those produced by experts, obtained from public databases used in the literature and private ones made available to the research group. The method offers contributions to the areas of Image Processing and Cardiology.
News published in Agência FAPESP Newsletter about the scholarship: