Evaluating the pre-processing impact on the generalization of deep learning networks for left ventricle segmentation

Autor(es)

Ribeiro, Matheus A. O. ; Nunes, Fatima L. S.

• Tipo de documento: Artigo Científico
• Fonte: 2021 IEEE INTERNATIONAL CONFERENCE ON BIOINFORMATICS AND BIOMEDICINE, BIBM; v. N/A, p. 8-pg., 2021-01-01.

Resumo

Deep learning networks have obtained promising results for left ventricle segmentation in magnetic resonance images. Training with datasets containing low diversity of examples impacts the generalization ability of these networks, which is a problem in the medical image context. Strategies such as data augmentation and post-processing techniques have been applied to improve learning and generalization, but the influence of pre-processing methods is still underexplored. This work aims to evaluate the impact of pre-processing steps on the generalization of deep learning networks for left ventricle segmentation. Experiments were conducted with public datasets considering different configurations of pre-processing steps. We also compare a novel approach for automatic region of interest (ROI) detection in relation to manual approaches. The results suggest that pre-processing procedures favor the generalization, even when other strategies are not present. New methods to detect the ROI with size invariability are promising to obtain adequate generalization ability in deep learning networks. (AU)

• Processo FAPESP: 19/22116-7 - Segmentação automática do ventrículo esquerdo em exames de ressonância magnética cardíaca
• Beneficiário: Matheus Alberto de Oliveira Ribeiro
• Modalidade de apoio: Bolsas no Brasil - Mestrado