AN END-END DEEP LEARNING FRAMEWORK FOR LUNG INFECTION RECOGNITION USING ATTENTION-BASED FEATURES AND CROSS AVERAGE POOLING

ABSTRACT

Automated detection of lung infections from medical imaging combined with computer vision has a great deal of promise for improving healthcare towards COVID-19 and its consequences due to restricted healthcare emergencies. Finding the affected tissues, segmenting them from lung images is difficult due to comparable neighboring tissues, hazy boundaries, and unpredictable infections. To overcome these issues, we propose a novel deep learning framework that employs attention-based feature vectors and cross average pooling to detect the lung infection from the images. Multimodal images, after enhancement are processed independently through a pretrained DenseNet where the feature extraction is performed from fully connected and average pooled layers. Instead of assigning equal weight to each feature value in the feature vectors, an attention weight is assigned to each feature to highlight how much attention should be paid to it. The attention-based features are then fused using cross average pooling to produce a discriminatory feature set leading to improved diagnosis. The fused features are passed through a deep learning (DL) modified neural network classifier to diagnose the infection. Experiments are performed on the standard Kaggle and Mendeley datasets containing 24,697 X-ray images and 8055 computed topography (CT) images. The results indicated an average accuracy of 99.2%, appreciable Kappa index of 98.11%, and F1 Score of 0.99. A one class accuracy of 99.5% is achieved for COVID-19. The proposed model is robust to noise when tested on degraded images. The results of our DL method for categorizing respiratory tract infections are compared to that of various existing DL models, demonstrating its effectiveness. © 2022 by Begell House, Inc.