Attention-Based Feature Extraction and Cross Average Pooling for Improved End-to-End Deep Learning Lung Infection Recognition

Abstract

Lung infections, such as pneumonia and COVID-19 among others, pose huge diagnostic challenges that can be tackled using sophisticated deep learning techniques. In this paper, we have presented the end-to-end deep learning framework by considering the attention-based feature extraction as well as cross average pooling methodology to improve lung infections recognition. Applications of traditional convolutional neural networks (CNNs) in medical image analysis show advantages; however, the features selected and represented by those models are far from optimal, thus leading to low accuracy as more complex cases (e. g., lung infections) emerge. To tackle this issue, we use an attention mechanism in our framework to concentrate on the informative parts of lung images which correspond to the infected regions with no noise. This attention-based feature extraction allows crucial information to be identified in training which will improve the models ability to recognize a person Moreover, a cross average pooling is applied to concatenate the feature maps, which can reflect richer and steadier infection patterns from chest X-ray images. Experimental results on publicly available lung infection datasets suggest that the proposed framework effectively improves over traditional CNN-based approaches, with higher accuracy, precision and recall rates. The results indicate that combining attentional mechanism and cross average pooling into deep learning architectures can significantly enhance the performance of automated detection and diagnose of lung infections, making it suitable for real clinical applications such as decision support auxiliary systems to assist in the diagnosis.