Progressive multi-scale attention neural network for pneumonia classification in chest X-rays

Abstract

We propose a novel Progressive Multi-Scale Attention Network (PMSAN) with an integrated Edge-Aware Loss function for improved pneumonia classification in chest X-rays. Unlike previous methods that overlook fine-grained edge information or fail to integrate multi-scale contextual features, our approach synergistically combines convolutional multi-scale feature extraction using depthwise separable convolutions with cross-layer feature fusion, Transformer blocks, advanced attention mechanisms, and a custom loss function that emphasizes diagnostically relevant edge details using Canny edge detection. Evaluated on the Kaggle chest X-ray pneumonia dataset—with optimal hyperparameters determined via extensive Optuna-based search—our model achieves a cross-validated accuracy of 97.3 % ± 0.4 % and an AUC of 0.995 ± 0.002 on the test set. Ablation studies and statistical significance tests confirm the contributions of each component, while visualizations demonstrate the model's ability to focus on clinically relevant regions. These substantial performance gains, along with a significant reduction in misdiagnoses, underscore the clinical potential of our efficient and accurate approach in supporting radiologists and improving patient outcomes.