PediaPulmoDx: Harnessing cutting edge preprocessing and explainable AI for pediatric chest X-ray classification with DenseNet121

Abstract

Pneumonia remains a predominant cause of morbidity and mortality among children, particularly in low- and middle-income countries, highlighting the urgent need for timely and accurate diagnostic approaches. Conventional diagnostic methods, however, are often hindered by limited healthcare infrastructure, subjective interpretation of Chest X-ray (CXR) images, and lack of awareness, particularly in underserved regions. While deep learning approaches show promise in automating pneumonia detection, they are often challenged by issues such as sensitivity to noisy images, class imbalance, and limited interpretability. In response to these challenges, we present pediapulmoDx, an innovative deep learning framework for pediatric pneumonia diagnosis from CXR images. This model integrates a suite of advanced preprocessing techniques, including Contrast-Limited Adaptive Histogram Equalization (CLAHE), Otsu's thresholding, data augmentation, and edge detection. Furthermore, pediapulmoDx employs robust feature extraction methods, such as Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG), to enhance feature representation and improve robustness against variations in image quality. At the core of our diagnostic framework is DenseNet121, a cutting-edge convolutional neural network (CNN) renowned for its superior feature extraction capabilities and classification efficiency. To enhance the transparency and interpretability of the model, we incorporate advanced explainable AI techniques, specifically Grad-CAM and Guided Grad-CAM, to generate heatmaps that highlight the pivotal image regions influencing the model's predictions. These visualizations not only provide clarity but also facilitate clinical validation by enabling practitioners to understand the rationale behind the models diagnostic outputs. Experimental evaluations reveal the outstanding performance of pediapulmoDx, achieving exceptional diagnostic metrics, including sensitivity of 99.60%, specificity of 99.80%, an F1-score of 99.70%, an AUC of 99.97%, and an overall accuracy of 99.97%. These results underscore the frameworks efficacy in delivering precise and reliable pediatric pneumonia detection. By combining state-of-the-art preprocessing techniques, robust feature extraction, advanced neural network architectures, and explainable AI, pediapulmoDx sets a new benchmark in pediatric pneumonia diagnosis, bridging the gap between cutting-edge technology and clinical decision-making. • Advanced Preprocessing : Applied GB, CLAHE, and Otsu's Thresholding for enhanced paediatric CXR quality. • Balanced Classification : Used data augmentation to handle class imbalance in pneumonia detection. • Robust Feature Extraction : Utilized LBP and HOG for texture and edge-based feature learning. • Optimized Model : Leveraged DenseNet121 for improved pneumonia classification. • Explainable AI : Integrated Grad-CAM for transparent decision-making.

Ähnliche Arbeiten