Efficient multi-class lung disease detection with SE-enhanced deep learning model: DenseLungNet-SE
TL;DR
Results suggest that coupling SE-guided channel recalibration with SVM-based decision-making yields a computationally practical and diagnostically reliable approach to automated multi-class pulmonary disease screening from chest radiographs.
Abstract
Pulmonary conditions including tuberculosis (TB), pneumonia, and coronavirus disease 2019 (COVID-19) pose considerable diagnostic difficulties owing to their overlapping visual presentations in chest radiographs, inconsistent imaging quality, and the growing strain placed on healthcare infrastructure. While automated detection systems driven by deep learning have demonstrated considerable potential, many such models carry high computational overhead and struggle to reliably distinguish between diseases with radiographically similar manifestations. This work presents DenseLungNet- squeeze-and-excitation (SE), a multi-class pulmonary disease classification system built on a hybrid architecture that merges DenseNet-121 with a SE attention block and a support vector machine (SVM) classifier. The SE block strengthens channel-level feature selectivity by directing the network's attention toward regions carrying the greatest diagnostic significance, while the SVM component provides robust decision boundaries in high-dimensional feature spaces. The system targets four categories of chest radiographs: COVID-19, TB, pneumonia, and healthy lungs. To enhance image consistency and prevent overfitting, a structured preprocessing and augmentation pipeline was incorporated into the training process. The system's performance was benchmarked against several established convolutional architectures using a comprehensive set of evaluation criteria. Beyond classification metrics, prediction calibration behavior and gradient-weighted class activation mapping (Grad-CAM) visualizations were examined to assess model transparency and clinical interpretability. Across repeated experimental runs, DenseLungNet-SE attained a mean accuracy of 95.05% ± 0.42 and an F1-score of 95.00% ± 0.45, with consistently balanced sensitivity and specificity values across all target classes. Memory footprint and inference latency were both lower than those recorded for the compared baseline architectures. These outcomes suggest that coupling SE-guided channel recalibration with SVM-based decision-making yields a computationally practical and diagnostically reliable approach to automated multi-class pulmonary disease screening from chest radiographs.