Hybrid Dimensionality Reduction and Ablation-Guided Optimization of Convolutional Neural Networks for Efficient Pulmonary Disease Classification from Chest X-Rays

Abstract:

Pulmonary disease classification by imaging with Chest X-Ray (CXR) is a vital aspect of proper and timely treatment planning and execution in patients with pulmonary disease. Given its accessibility and affordability, it has broad applicability in a healthcare setting. Although deep learning techniques have progressed, specifically in the area of Convolutional Neural Network (CNN) technology, real-world implementation of CXR-based classifiers in pulmonary disease classification is interfered with by problems such as the minute diagnostic patterns in diseases and extreme class imbalance, among other factors related to its training and applicability in CXR-based imaging analysis in a healthcare setting like a hospital with scarce resources. The proposed improved solution to this problem involves advanced CNN technology with additional specifics in its deep learning analysis. This method leverages two base models that have demonstrated efficiency in medical image analysis, namely DenseNet121 and ResNet50. The hybrid PCA-Autoencoder is used to compress the feature space through dimensionality reduction while minimizing loss of diagnostic information. Ablation-guided tuning methodically removes unnecessary layers and parameters to enhance efficiency. Using 5-fold stratified cross-validation, the optimized architecture is trained and assessed on four sizable publicly accessible CXR datasets: NIH ChestX-ray14, CheXpert, RSNA Pneumonia, and TBX11K. Compared to the baseline, the improved DenseNet121 reduced the inference time by 40.1% and the model size by 36.6% while achieving 97.9% accuracy, 97.5% precision, 98.2% recall, and an AUC of 0.986. The results are indicative that the hybrid optimization method proposed herein improves the classification performance while overcoming scalability and computing challenges; thus, it presents a deployable solution for real-time detection of pulmonary illness in diverse clinical settings.

Keywords: Pulmonary disease classification, Chest X-ray, Convolutional neural network, Dimensionality reduction, Ablation-guided parameter tuning, Computational efficiency.