Karim Faizan, Shah Munam Ali, Khattak Hasan Ali, Ameer Zoobia, Shoaib Umar, Rauf Hafiz Tayyab, Al-Turjman Fadi
COMSATS University Islamabad, Pakistan.
National University of Sciences and Technology (NUST), Islamabad, Pakistan.
Appl Soft Comput. 2022 Jul;124:109077. doi: 10.1016/j.asoc.2022.109077. Epub 2022 May 31.
Machine Learning and computer vision have been the frontiers of the war against the COVID-19 Pandemic. Radiology has vastly improved the diagnosis of diseases, especially lung diseases, through the early assessment of key disease factors. Chest X-rays have thus become among the commonly used radiological tests to detect and diagnose many lung diseases. However, the discovery of lung disease through X-rays is a significantly challenging task depending on the availability of skilled radiologists. There has been a recent increase in attention to the design of Convolution Neural Networks (CNN) models for lung disease classification. A considerable amount of training dataset is required for CNN to work, but the problem is that it cannot handle translation and rotation correctly as input. The recently proposed Capsule Networks (referred to as CapsNets) are new automated learning architecture that aims to overcome the shortcomings in CNN. CapsNets are vital for rotation and complex translation. They require much less training information, which applies to the processing of data sets from medical images, including radiological images of the chest X-rays. In this research, the adoption and integration of CapsNets into the problem of chest X-ray classification have been explored. The aim is to design a deep model using CapsNet that increases the accuracy of the classification problem involved. We have used convolution blocks that take input images and generate convolution layers used as input to capsule block. There are 12 capsule layers operated, and the output of each capsule is used as an input to the next convolution block. The process is repeated for all blocks. The experimental results show that the proposed architecture yields better results when compared with the existing CNN techniques by achieving a better area under the curve (AUC) average. Furthermore, DNet checks the best performance in the ChestXray-14 data set on traditional CNN, and it is validated that DNet performs better with a higher level of total depth.
机器学习和计算机视觉一直处于抗击新冠疫情的前沿。放射学通过对关键疾病因素的早期评估，极大地改善了疾病诊断，尤其是肺部疾病的诊断。胸部X光检查因此成为检测和诊断多种肺部疾病常用的放射学检查手段之一。然而，依靠X光发现肺部疾病是一项极具挑战性的任务，这取决于是否有技术娴熟的放射科医生。最近，人们越来越关注用于肺部疾病分类的卷积神经网络（CNN）模型的设计。CNN运行需要大量的训练数据集，但问题在于它无法正确处理输入的平移和旋转。最近提出的胶囊网络（CapsNets）是一种新的自动学习架构，旨在克服CNN的缺点。胶囊网络对于旋转和复杂平移至关重要。它们所需的训练信息要少得多，适用于处理来自医学图像的数据集，包括胸部X光的放射图像。在本研究中，探索了将胶囊网络应用于胸部X光分类问题并进行整合。目的是使用胶囊网络设计一个深度模型，以提高相关分类问题的准确性。我们使用了卷积块，其接收输入图像并生成用作胶囊块输入的卷积层。有12个胶囊层运行，每个胶囊的输出用作下一个卷积块的输入。对所有块重复此过程。实验结果表明，与现有的CNN技术相比，所提出的架构通过获得更好的曲线下面积（AUC）平均值产生了更好的结果。此外，DNet在传统CNN的ChestXray - 14数据集中表现出最佳性能，并且验证了DNet在总深度更高的情况下表现更好。
