Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany.
International Max Planck Research School for Physics of Light, Erlangen, Germany.
Int J Comput Assist Radiol Surg. 2021 Jun;16(6):967-978. doi: 10.1007/s11548-021-02340-1. Epub 2021 Apr 30.
With the recent development of deep learning technologies, various neural networks have been proposed for fundus retinal vessel segmentation. Among them, the U-Net is regarded as one of the most successful architectures. In this work, we start with simplification of the U-Net, and explore the performance of few-parameter networks on this task.
We firstly modify the model with popular functional blocks and additional resolution levels, then we switch to exploring the limits for compression of the network architecture. Experiments are designed to simplify the network structure, decrease the number of trainable parameters, and reduce the amount of training data. Performance evaluation is carried out on four public databases, namely DRIVE, STARE, HRF and CHASE_DB1. In addition, the generalization ability of the few-parameter networks are compared against the state-of-the-art segmentation network.
We demonstrate that the additive variants do not significantly improve the segmentation performance. The performance of the models are not severely harmed unless they are harshly degenerated: one level, or one filter in the input convolutional layer, or trained with one image. We also demonstrate that few-parameter networks have strong generalization ability.
It is counter-intuitive that the U-Net produces reasonably good segmentation predictions until reaching the mentioned limits. Our work has two main contributions. On the one hand, the importance of different elements of the U-Net is evaluated, and the minimal U-Net which is capable of the task is presented. On the other hand, our work demonstrates that retinal vessel segmentation can be tackled by surprisingly simple configurations of U-Net reaching almost state-of-the-art performance. We also show that the simple configurations have better generalization ability than state-of-the-art models with high model complexity. These observations seem to be in contradiction to the current trend of continued increase in model complexity and capacity for the task under consideration.
随着深度学习技术的发展，各种神经网络被提出用于眼底视网膜血管分割。其中，U-Net 被认为是最成功的架构之一。在这项工作中，我们从 U-Net 的简化开始，探索这个任务中少量参数网络的性能。
我们首先使用流行的功能块和附加分辨率层来修改模型，然后转而探索网络架构的压缩极限。实验旨在简化网络结构、减少可训练参数数量和减少训练数据量。性能评估在四个公共数据库上进行，即 DRIVE、STARE、HRF 和 CHASE_DB1。此外，还比较了少量参数网络的泛化能力与最先进的分割网络。
我们证明了添加的变体并没有显著提高分割性能。除非模型严重退化：一个级别、输入卷积层中的一个滤波器或只用一张图像进行训练，否则模型的性能不会受到严重影响。我们还证明了少量参数网络具有很强的泛化能力。
直到达到所述限制，U-Net 才产生相当好的分割预测，这是违反直觉的。我们的工作有两个主要贡献。一方面，评估了 U-Net 的不同元素的重要性，并提出了能够完成任务的最小 U-Net。另一方面，我们的工作表明，通过非常简单的 U-Net 配置可以解决视网膜血管分割问题，达到几乎最先进的性能。我们还表明，简单的配置比具有高模型复杂度的最先进模型具有更好的泛化能力。这些观察结果似乎与考虑到的任务中模型复杂性和容量持续增加的当前趋势相矛盾。
