Purpose

The robust and automatic segmentation of the pulmonary lobe is vital to surgical planning and regional image analysis of pulmonary related diseases in real-time Computer Aided Diagnosis systems. While a number of studies have examined this issue, the segmentation of unclear borders of the five lobes of the lung remains challenging because of incomplete fissures, the diversity of anatomical pulmonary information, and obstructive lesions caused by pulmonary diseases. This study proposes a model called Regularized Pulmonary Lobe Segmentation Network to accurately predict the lobes as well as the borders.

Methods

First, a 3D fully convolutional network is constructed to extract contextual features from computed tomography images. Second, multi-task learning is employed to learn the segmentations of the lobes and the borders between them to train the neural network to better predict the borders via shared representation. Third, a 3D depth-wise separable de-convolution block is proposed for deep supervision to efficiently train the network. We also propose a hybrid loss function by combining cross-entropy loss with focal loss using adaptive parameters to focus on the tissues and the borders of the lobes.

Results

Experiments are conducted on a dataset annotated by experienced clinical radiologists. A 4-fold cross-validation result demonstrates that the proposed approach can achieve a mean dice coefficient of 0.9421 and average symmetric surface distance of 1.3546 mm, which is comparable to state of the art methods. The proposed approach has the capability to accurately segment voxels that are near the lung wall and fissure.

Conclusion

In this paper, a 3D fully convolutional networks framework is proposed to segment pulmonary lobes in chest CT images accurately. Experimental results show the effectiveness of the proposed approach in segmenting the tissues as well as the borders of the lobes.

中文翻译：

---

RPLS-Net：基于3D全卷积网络和多任务学习的肺叶分割

目的

在实时计算机辅助诊断系统中，肺叶的强大而自动的分割对于肺部相关疾病的手术计划和区域图像分析至关重要。尽管许多研究已经对此问题进行了研究，但由于裂孔不完整，肺部解剖学信息的多样性以及由肺部疾病引起的阻塞性病变，对肺部五个肺叶边界不清晰的分割仍然具有挑战性。这项研究提出了一种称为正则化肺叶分割网络的模型，以准确预测肺叶和边界。

方法

首先，构建3D全卷积网络以从计算机断层扫描图像中提取上下文特征。其次，采用多任务学习来学习叶的分割及其之间的边界，以训练神经网络通过共享表示更好地预测边界。第三，提出了一种3D深度可分离的反卷积块，用于深度监控，以有效地训练网络。我们还提出了一种混合损失函数，该方法通过使用自适应参数将交叉熵损失与焦点损失结合起来，以关注于组织和叶的边界。

结果

实验是在经验丰富的临床放射科医生注释的数据集上进行的。4倍交叉验证结果表明，该方法可以实现0.9421的平均骰子系数和1.3546 mm的平均对称表面距离，与现有技术水平相当。所提出的方法具有准确分割靠近肺壁和裂隙的体素的能力。

结论

在本文中，提出了一个3D全卷积网络框架以准确分割胸部CT图像中的肺叶。实验结果表明，该方法在分割组织以及肺叶边界方面是有效的。