Does Isotropy Improve 3D Deep Learning Segmentation? A Comparative Study on Cardiac MRI

• Matheus A. O. Ribeiro USP
• Marco A. Gutierrez USP
• Fátima L. S. Nunes USP

Resumo

Deep learning networks have excelled in several fields, achieving state-of-the-art quality in many complex problems. In the Medical Image Processing field, various three-dimensional (3D) deep learning methods have been proposed and validated for automatic segmentation. One of the main advantages of 3D strategies over simpler two-dimensional (2D) approaches is improved consistency by leveraging inter-slice contextual information. However, comparative studies show that 3D and 2D approaches achieve similar results in many segmentation tasks. A common property that affects mainly 3D strategies and is often ignored by many network designs is the data anisotropy. While very common in several medical imaging modalities, anisotropy is often disregarded by many studies, and its effect on 3D performance is still an underexplored topic. Considering this limitation, in this work, we investigate the advantages of promoting data isotropy on 3D deep learning segmentation. To do this, we propose a fully automated 3D segmentation pipeline that explores data standardization and slice interpolation methods to achieve adjustable levels of anisotropy. We performed cross-validation experiments considering two public cardiac MRI datasets for left ventricle segmentation, which is a popular problem containing highly anisotropic data. Then, we compare the performance of 2D and 3D U-nets trained on different anisotropy levels. The results indicate that promoting isotropy can significantly improve the performance and generalization ability of 3D deep learning approaches. However, the high computational cost, aligned with low improvements when compared to 2D strategies, may discourage its applicability in clinical practice.

Palavras-chave: Deep learning, Image segmentation, Interpolation, Three-dimensional displays, Anisotropic magnetoresistance, Magnetic resonance imaging, Two-dimensional displays, Pipelines, Anisotropic, Computational efficiency