Long-range decoder skip connections: Exploiting multi-context information for cardiac image segmentation

• Autores: Gutierrez-Castilla N., Da Torres R., Falcao A.X., Kozerke S., Schwitter J., Masci P.-G., Montoya-Zegarra J.A.
• Formato: artículo
• Fecha de Publicación: 2019
• Institución: Consejo Nacional de Ciencia Tecnología e Innovación
• Repositorio: CONCYTEC-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:repositorio.concytec.gob.pe:20.500.12390/2697
• Enlace del recurso: https://hdl.handle.net/20.500.12390/2697; https://doi.org/10.1109/SIBGRAPI.2019.00017
• Nivel de acceso: acceso abierto
• Materia: Semantic image segmentation; Biomedical imaging; Cardiac image analysis; Deep learning; http://purl.org/pe-repo/ocde/ford#2.02.03

The heart is one of the most important organs in our body and many critical diseases are associated with its malfunctioning. To assess the risk for heart diseases, Magnetic Resonance Imaging (MRI) has become the golden standard imaging technique, as it provides to the clinicians stacks of images for analyzing the heart structures, such as the ventricles, and thus to make a diagnosis of the patient's health. The problem is that examination of these stacks, often based on the delineation of heart structures, is tedious and error prone due to inter-and intra-variability among manual delineations. For this reason, the investigation of fully automated methods to support heart segmentation is paramount. Most of the successful methods proposed to solve this problem are based on deep-learning solutions. Especially, encoder-decoder architectures, such as the U-Net [1], have demonstrated to be very effective architectures for medical image segmentation. In this paper, we propose to use long-range skip connections on the decoder-part to incorporate multi-context information onto the predicted segmentation masks and also to improve the generalization of the models. In addition, our method obtains smoother segmentations through the combination of feature maps from different stages onto the final prediction layer. We evaluate our approach in the ACDC [2] and LVSC [3] heart segmentation challenges. Experiments performed on both datasets demonstrate that our approach leads to an improvement on both the total Dice score and the Ejection Fraction Correlation, when combined with state-of-the-art encoder-decoder architectures.