Accurate extraction of semantic objects such as ventricles and myocardium from magnetic resonance (MR) images is one essential but very challenging task for the diagnosis of the cardiac diseases. To tackle this problem, in this paper, an automatic end-to-end supervised deep learning framework is proposed, using a multi-task learning based U-Net (MTL-UNet). Specifically, an edge extraction module and a fusion-based module are introduced for effectively capturing the contextual information such as continuous edges and consistent spatial patterns in terms of intensity and texture features. With a weighted triple loss including the dice loss, the cross-entropy loss and the edge loss, the accuracy of object segmentation and extraction has been effectively improved. Extensive experiments on the publicly available ACDC 2017 dataset have validated the efficacy and efficiency of the proposed MTL-UNet model.

从磁共振 (MR) 图像中准确提取语义对象，例如心室和心肌，是诊断心脏病的一项必不可少但非常具有挑战性的任务。为了解决这个问题，在本文中，使用基于多任务学习的 U-Net (MTL-UNet) 提出了一种自动端到端监督深度学习框架。具体来说，引入了边缘提取模块和基于融合的模块，以在强度和纹理特征方面有效地捕获上下文信息，例如连续边缘和一致的空间模式。通过包含骰子损失、交叉熵损失和边缘损失的加权三重损失，有效提高了对象分割和提取的准确性。