Efficient and explainable pneumoconiosis screening from chest radiographs using a task-optimized CNN

Abstract

Context—Pneumoconiosis is a chronic occupational lung disease characterized by subtle, low-contrast, and diffusely distributed parenchymal abnormalities on chest radiographs. These imaging characteristics, combined with limited dataset availability and severe class imbalance, make reliable automated detection challenging. Although recent deep learning approaches have shown promising results, they often rely on increasingly complex and computationally demanding architectures, while the alignment between network design and disease-specific radiographic patterns remains insufficiently explored.Objective—This study aims to develop an efficient and task-specific deep learning framework for the automatic detection of pneumoconiosis from posteroanterior chest radiographs. The primary objective is to investigate whether a compact and optimized convolutional architecture can achieve competitive performance without relying on excessive model complexity, while maintaining high sensitivity required for screening applications.Method—A lightweight convolutional neural network, termed ResPneumoNet, is proposed, where network depth and convolutional kernel size are treated as task-dependent architectural hyperparameters. These parameters are jointly optimized using Bayesian hyperparameter optimization to align the receptive field and feature extraction capacity with pneumoconiosis-specific radiographic patterns. To address class imbalance, diffusion-based synthetic image generation (MedFusion) is employed to construct a balanced dataset while preserving anatomical and textural consistency. The model is trained and evaluated on chest radiographs resized to 224×224 pixels, using standard classification metrics and ROC analysis. Additionally, lung-region–constrained Grad-CAM is utilized to assess model interpretability.Results—On the independent real test set consisting of 112 normal and 28 pneumoconiosis-positive radiographs, ResPneumoNet achieved an accuracy of 76.43%, sensitivity of 64.29%, specificity of 79.46%, and an AUC of 0.752, with 107.1k trainable parameters. Deeper reference models achieved higher overall discriminative performance, indicating that ResPneumoNet prioritizes architectural compactness over maximum classification accuracy. Lung-region–constrained Grad-CAM analyses showed that model attention was generally concentrated within lung-field regions, providing qualitative support for anatomically relevant decision behavior.Conclusion—The findings indicate that task-aligned architectural optimization can provide a compact and computationally efficient alternative for pneumoconiosis detection. Although the proposed framework did not outperform deeper reference models in overall discrimination, it offered reduced model complexity with screening-oriented sensitivity. Future work will focus on external validation across diverse datasets and extending the approach toward severity-aware analysis.