CXRNet: CNN-attention based CXR image classifier

Abstract

Chest X-ray (CXR) images are widely accepted for the diagnosis of lung diseases. The X-ray machinery is widely available but the number of radiologists who interpret these images is very limited. Therefore, the development of an automated disease classification system is a need for the healthcare industry. The existing testing methods take hours to days to generate the testing result and have low detection accuracy and high false detection rate. Furthermore, the testing kits are costly, and availability is limited. Therefore, Convolutional Neural Network (CNN) based framework is proposed to address these limitations. Four pre-trained frameworks, ResNet50V2, InceptionV3, NASNetMobile, and Xception are used to generate highly dense features, refine the features using the attention module, and then fused these features to classify the diseases using CXR images. In addition, extensive experiments were carried out on the activation functions (Relu, Leaky Relu, and Tanh), which help to improve the results. Evaluation is done on the three enriched CXR datasets DS-1, DS-2, and DS-3 to examine the performance of the proposed framework in terms of binary classification and multi-class classification. The proposed model achieved a class-5 accuracy of 92.89% on dataset DS-3 and the class-3 classification accuracy of 95.22%, 92.25% on dataset DS-1 and DS-2, respectively. The class-2 classification accuracy on dataset DS-1 and DS-2 is 99.23% and 98.85%, respectively.