An Improving Lung Disease Detection by Combining Ensemble Deep Learning and Maximum Mean Discrepancy Transfer Learning
Year: 2024
Type: article
Abstract: In accordance to the World Health Organization (WHO), various pulmonary diseases cause thousands of deaths annually.The early diagnosis is required to lessen the mortality rate.For this reason, A Convolutional Neural Network (CNN)-based Lung Disease (LD) detection system is developed to classify segregated lung sections into various pulmonary diseases types.However, epistemic uncertainty in the scanned images affecting the performance of detection classifiers.Hence, in this paper, a multi-modal approach is proposed to solve the epistemic uncertainty issue and provides a reliable solution for rapid detection of various LD types from CXR images.In this method, CT images are additionally used to improve model's performance as it contains detailed information that might be exploited to provide efficient results.Initially, the collected images are segmented using U-Net model to get enhanced lung Region of Interest (ROIs).Then ResNet50, DenseNet121, InceptionResNetV2 and XceptionV3 are used to hierarchically extract informative and discriminative features from collected CXR and CT images.The retrieved deep features are fed into the Ensemble-Convolutional Long Short Term Memory with Extreme Machine Learning (E-conLSTM-ELM) to minimize the computational time and increase the accuracy.Moreover, Transfer Learning (TL) model is employed to learn the weight of the E-conLSTM-ELM to exchange the knowledge between features and classes relation among CXR and CT images.Also, the domain adaptation approach is a variant of TL model that relies on employing similar datasets for a shared learning problem.This adaption strategy reduces the domain shift (data dispersion) using Maximum Mean Discrepancy (MMD).The shared semantic features from CT images through TL improve the in-depth learning of softmax layer to classify different LD types.The proposed work is simply named as Convolutional LD Scan (CovLscan) framework The test outcomes reveal that the CovLscan model accomplishes an overall accuracy of 95.46% and 96.15% on the collected ChestX-ray8 and NIH-CXR datasets, which is higher than the existing models like Automated Hierarchical Deep Learning based LD Diagnosis(AHDL-LDD), EfficientNet version2-Medium (EfficientNet v2-M), Lung diseases prediction Network22(LungNet22), Chest tract disorder prediction using Dilated Convolutional Network(CDCNet) and Auction-Based Optimization Algorithm-CNN (ABOA-CNN).
Language: English
Topic: COVID-19 diagnosis using AI
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4401702477