Reducing False-Negative Risk in Automated Pneumonia Screening Using Attention-Guided Deep Learning

Abstract

Pneumonia continues to be a major source of morbidity and mortality worldwide, especially in children, the elderly, and people with impaired immune systems. Due to its low cost and widespread availability, chest imaging radiology is the most often utilised diagnostic modality; yet, accurate interpretation is difficult and heavily dependent on radiologist expertise, which can result in clinically significant missed diagnosis. While deep learning-based methods have demonstrated potential for automated pneumonia detection, many of the models now in use rely on global feature learning, have poor interpretability, and do not adequately address the danger of false negatives. An attention-enhanced deep learning framework for clinically accurate pneumonia identification from chest imaging radiology is proposed in this study.To enhance spatial feature representation and highlight diagnostically significant lung regions, the framework combines a self-attention mechanism with a pretrained VGG16 backbone. To guarantee stable optimisation and efficient task adaptability, a two-stage training approach is used. The suggested model is tested using a single experimental methodology against many cutting-edge convolutional neural network architectures. The suggested framework obtains an accuracy of 96.3%, recall of 98.0%, F1-score of 97.5%, and ROC–AUC of 0.972, according to experimental results. Significantly, compared to the baseline VGG16 model, false-negative predictions are decreased from 27 to 17.