Sequential and ensemble hybrid approaches for left ventricle segmentation in cardiac MR: A systematic mapping

Segmentation of the left ventricle in short-axis cardiac magnetic resonance images is an important step to assess heart function. Given the considerable human effort required for segmenting an entire exam, several automatic and semi-automatic methods have been proposed, following five main segmentation categories: Atlas, Graphs, Deformable Models, Artificial Intelligence, and Image-based strategies. In recent years, Artificial Intelligence, especially Deep Learning networks, have presented performance very similar to experts considering commonly used evaluation metrics. However, state-of-the-art methods still have limitations, such as the need for huge amounts of diverse instances for training, production of segmentation with anatomical errors, overfitting, low explainability, and generalization problems. An alternative to reduce these limitations and improve performance is to develop hybrid approaches that combine methods from different segmentation categories, exploiting their individual advantages. Considering that the development of hybrid approaches is still underexplored, in this work, we present a review of hybrid approaches for left ventricle segmentation in cardiac magnetic resonance images. After categorizing single techniques from 35 included studies as image-based, atlas, graphs, deformable models, and artificial intelligence, we propose a new classification of hybrid methods as sequential and ensemble, and analyze their main characteristics. We highlight the advantages and challenges of each hybrid combination and discuss the limitations of current evaluation procedures. The results indicate that hybrid approaches can help to reduce present limitations, but the lack of adequate evaluation metrics, generalization, and statistical significance analysis hinders the assessment of clinical applicability and reliability.