da Silva Matheus V, Ouellette Julie, Lacoste Baptiste, Comin Cesar H
Department of Computer Science, Federal University of São Carlos, São Carlos, SP, Brazil.
Department of Cellular and Molecular Medicine, Faculty of Medicine, University of Ottawa, Ottawa, ON, Canada; Neuroscience Program, Ottawa Hospital Research Institute, Ottawa, ON, Canada.
Comput Methods Programs Biomed. 2022 Oct;225:107021. doi: 10.1016/j.cmpb.2022.107021. Epub 2022 Jul 16.
Convolutional Neural Networks (CNNs) can provide excellent results regarding the segmentation of blood vessels. One important aspect of CNNs is that they can be trained on large amounts of data and then be made available, for instance, in image processing software. The pre-trained CNNs can then be easily applied in downstream blood vessel characterization tasks, such as the calculation of the length, tortuosity, or caliber of the blood vessels. Yet, it is still unclear if pre-trained CNNs can provide robust, unbiased, results in downstream tasks involving the morphological analysis of blood vessels. Here, we focus on measuring the tortuosity of blood vessels and investigate to which extent CNNs may provide biased tortuosity values even after fine-tuning the network to a new dataset under study.
We develop a procedure for quantifying the influence of CNN pre-training in downstream analyses involving the measurement of morphological properties of blood vessels. Using the methodology, we compare the performance of CNNs that were trained on images containing blood vessels having high tortuosity with CNNs that were trained on blood vessels with low tortuosity and fine-tuned on blood vessels with high tortuosity. The opposite situation is also investigated.
We show that the tortuosity values obtained by a CNN trained from scratch on a dataset may not agree with those obtained by a fine-tuned network that was pre-trained on a dataset having different tortuosity statistics. In addition, we show that improving the segmentation accuracy does not necessarily lead to better tortuosity estimation. To mitigate the aforementioned issues, we propose the application of data augmentation techniques even in situations where they do not improve segmentation performance. For instance, we found that the application of elastic transformations was enough to prevent an underestimation of 8% of blood vessel tortuosity when applying CNNs to different datasets.
The results highlight the importance of developing new methodologies for training CNNs with the specific goal of reducing the error of morphological measurements, as opposed to the traditional approach of using segmentation accuracy as a proxy metric for performance evaluation.
卷积神经网络（CNN）在血管分割方面能提供出色的结果。CNN的一个重要方面是它们可以在大量数据上进行训练，然后例如在图像处理软件中可用。预训练的CNN随后可以很容易地应用于下游血管特征描述任务，如血管长度、曲折度或管径的计算。然而，在涉及血管形态分析的下游任务中，预训练的CNN是否能提供可靠、无偏差的结果仍不清楚。在这里，我们专注于测量血管的曲折度，并研究即使在将网络微调至新的研究数据集后，CNN在多大程度上可能提供有偏差的曲折度值。
我们开发了一种程序，用于量化CNN预训练在涉及血管形态特性测量的下游分析中的影响。使用该方法，我们比较了在包含高曲折度血管的图像上训练的CNN与在低曲折度血管上训练并在高曲折度血管上微调的CNN的性能。也研究了相反的情况。
我们表明，在一个数据集上从头开始训练的CNN获得的曲折度值可能与在具有不同曲折度统计数据的数据集上预训练后微调的网络获得的曲折度值不一致。此外，我们表明提高分割精度不一定会导致更好的曲折度估计。为了减轻上述问题，我们建议即使在数据增强技术不能提高分割性能的情况下也应用它们。例如，我们发现当将CNN应用于不同数据集时，应用弹性变换足以防止低估8%的血管曲折度。
结果强调了开发新方法来训练CNN的重要性，其具体目标是减少形态测量的误差，而不是使用分割精度作为性能评估的替代指标的传统方法。
