Deep transfer learning approaches for bleeding detection in endoscopy images
Deep transfer learning approaches for bleeding detection in endoscopy images
复制标题
DOI：
10.1016/j.compmedimag.2020.101852
复制
发表时间：
2021-01-23
影响因子：
5.7
通讯作者：
Siciliano, Pietro
中科院分区：
文献类型：
--
作者：
Caroppo, Andrea;Leone, Alessandro;Siciliano, Pietro
无线胶囊式内窥镜是一种非侵入性的无线成像工具，在过去几年中发展迅速。使用这项技术的主要限制因素之一是，它会产生大量的图像，由医生进行分析是一个非常耗时的过程。在这一研究领域，这一问题的管理已经通过计算机辅助诊断系统的开发得到解决，由于该系统，胶囊采集的图像的自动检查和分析得到了明显的改善。最近，随着深度学习方法的出现，内窥镜图像分类取得了很大的进步。所提出的专家系统采用三个预训练的深度卷积神经网络进行特征提取。为了构建有效的特征集，从VGG19，InceptionV3和ResNet50模型中选择特征，并使用最小冗余最大相关性方法和不同的融合规则进行融合。最后，监督机器学习算法进行分类的图像使用提取的功能分为两类：出血和非出血图像。为了进行性能评估，在两个标准基准数据集上进行了一系列实验。已经观察到，所提出的架构优于单个深度学习架构，在考虑三种不同融合规则的已知最先进数据集上检测出血区域的平均准确度为97.65%和95.70%，使用均值池作为融合规则和支持向量机作为分类器获得准确度和训练时间方面的最佳组合。
Wireless capsule endoscopy is a non-invasive, wireless imaging tool that has developed rapidly over the last several years. One of the main limiting factors using this technology is that it produces a huge number of images, whose analysis, to be done by a doctor, is an extremely time-consuming process. In this research area, the management of this problem has been addressed with the development of Computer-aided Diagnosis systems thanks to which the automatic inspection and analysis of images acquired by the capsule has clearly improved. Recently, a big advance in classification of endoscopic images is achieved with the emergence of deep learning methods. The proposed expert system employs three pre-trained deep convolutional neural networks for feature extraction. In order to construct efficient feature sets, the features from VGG19, InceptionV3 and ResNet50 models are then selected and fused using the minimum Redundancy Maximum Relevance method and different fusion rules. Finally, supervised machine learning algorithms are employed to classify the images using the extracted features into two categories: bleeding and nonbleeding images. For performance evaluation a series of experiments are performed on two standard benchmark datasets. It has been observed that the proposed architecture outclass the single deep learning architectures, with an average accuracy in detection bleeding regions of 97.65 % and 95.70 % on well-known state-of-the-art datasets considering three different fusion rules, with the best combination in terms of accuracy and training time obtained using mean value pooling as fusion rule and Support Vector Machine as classifier.