Zhuo Zhongshuo, Huang Jianping, Lu Ke, Pan Daru, Feng Shouting
School of Physics and Telecommunication Engineering, South China Normal University, Guangzhou 510006, China.
University of Chinese Academy of Sciences, No. 19A Yuquan Road, Beijing 100049, China.
Comput Methods Programs Biomed. 2020 Nov;196:105508. doi: 10.1016/j.cmpb.2020.105508. Epub 2020 May 31.
Retinal vessel segmentation (RVS) helps in diagnosing diseases such as hypertension, cardiovascular diseases, and others. Convolutional neural networks are widely used in RVS tasks. However, how to comprehensively evaluate the segmentation results and how to improve the networks' learning ability are two great challenges.
In this paper, we proposed an ingenious index: fusion score (FS), which provides an overall measure for those binary images. The FS converts multiple metrics into a single target, and therefore facilitates the optimal threshold's selection and models' comparison. In addition, We simultaneously combined size-invariant feature maps and dense connectivity together to improve the traditional CNN's learning ability. Therefore, a size-invariant convolutional network with dense connectivity is designed for RVS. The size-invariant skill helps the deep layers create feature maps with high resolution. The dense connectivity technique is utilized to integrate those hierarchical features and reuse characteristic maps to enhance the network's learning ability. Finally, an optimized threshold is used on the output image to obtain a binary image.
The results of experiments conducted on two shared retinal image databases, DRIVE and STARE, demonstrate that our approach outperforms other techniques when evaluated in terms of F1-score, Matthews correlation coefficient (MCC), G-mean and FS. In addition, the cross training reveals that our method has stronger robustness with respect to training sets. Segmenting a 565 × 584 image only takes 39 ms with a single GPU (graphics processing unit).
Compared with those traditional metrics, the FS is a better indicator to measure the results of RVS tasks. The experimental results revealed that the proposed method is more suitable for real-world applications.
视网膜血管分割（RVS）有助于诊断高血压、心血管疾病等病症。卷积神经网络在RVS任务中被广泛应用。然而，如何全面评估分割结果以及如何提高网络的学习能力是两大挑战。
在本文中，我们提出了一个巧妙的指标：融合分数（FS），它为那些二值图像提供了一个整体度量。FS将多个指标转换为单个目标，因此便于选择最优阈值和比较模型。此外，我们同时将尺寸不变特征图和密集连接结合在一起，以提高传统卷积神经网络的学习能力。因此，设计了一种具有密集连接的尺寸不变卷积网络用于RVS。尺寸不变技术有助于深层创建高分辨率的特征图。密集连接技术用于整合那些分层特征并重用特征图以增强网络的学习能力。最后，对输出图像使用优化后的阈值以获得二值图像。
在两个共享的视网膜图像数据库DRIVE和STARE上进行的实验结果表明，在F1分数、马修斯相关系数（MCC）、G均值和FS方面进行评估时，我们的方法优于其他技术。此外，交叉训练表明我们的方法在训练集方面具有更强的鲁棒性。使用单个图形处理单元（GPU）分割一幅565×584的图像仅需39毫秒。
与那些传统指标相比，FS是衡量RVS任务结果的更好指标。实验结果表明所提出的方法更适合实际应用。
