Abstract
Lung disease is common throughout the world. These include chronic obstructive pulmonary disease, pneumonia, asthma, tuberculosis, fibrosis, etc. Timely diagnosis of lung disease is essential. Many image processing and machine learning (ML) models have been developed to make more accurate with less time and effort. But still, the degraded accuracy performance was resulted in these systems. In recent days, Deep Learning (DL) model plays an important role in classifying the lung diseases for the early prediction process. Amongst, an efficient Convolutional Neural Network (CNN)-based lung disease detection system is developed with additional layers to classify the segregated lung sections into various lung diseases using Chest X-Ray (CXR) images. However, this model results in epistemic uncertainty which effects the performance of DL models employed for lung disease diagnosis. Hence, in this paper, a multi-modal approach called Ensemble Deep Lung Disease Predictor (EDepLDP) framework is proposed to solve the epistemic uncertainty issue and develops a reliable solution for rapid detection of various diseases using CXR and Computerized Tomography (CT) images. Initially, the collected images are segmented using U-Net model to get enhanced lung Region of Interest (ROIs). Then, InceptionResNetV2 and Xception are used to hierarchically extract informative features from segmented CXR images and discriminative features from segmented CT images respectively. The extracted deep features are fed into the softmax layer of conGRU-LSTM to perform the classification task. Moreover, the TL model is developed to learn the weight for the InceptionResNetV2, Xception and conGRU-LSTM which is obtained from the pre-trained Efficient-Net model. Also, the domain adaptation strategy is a subset of TL model which mainly addresses the situation where different but related datasets for a common learning task. This adaption strategy reduces the domain shift or data distribution using Maximum Mean Discrepancy (MMD) for the efficient classification of various lung diseases. The test outcomes reveal that the EDepLDP model accomplishes an overall accuracy of 92% and 92% on the collected CXR and CT images which is contrasted with the classical CNN models.
Keyword
Chest X-Rays, Computerized Tomography, Epistemic Uncertainty, Transfer Learning, Normalized Feature Inputs
PDF Download (click here)
|