Deep Learning for Reliable Healthcare: Insights from Chest X-Ray Prediction

Abstract

Pneumonia is a principal ailment causing morbidity and mortality in Bangladesh, particularly in children with inequities in health care access, which motivated this study. Therefore, this study presents a hybrid deep learning model that employs two separate approaches - Xception and DenseNet121, to enable pneumonia detection from chest X-ray images in a highly automated fashion - using a dataset from Mendeley (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{3, 2 5 6}$</tex> images: 504 normal, 985 pneumonia). Image preprocessing (cropping, resizing to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$180 \times 180$</tex> pixels, normalization, and augmentation) strategies were engaged, and 5-fold cross-validation provided robustness, producing a training accuracy of 98%, validation accuracy greater than 97%, and a mean accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$95.71 \% \pm 1.76 {\%}$</tex> (loss: <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.30 \pm 0.21$</tex>) greater than Inception v3, DenseNet121, and MobileNetV2 comparative baseline methods. Saliency maps provide further explainability to patients and their caregivers, giving clinical trust in the methods. Additional analytic evaluations used confusion matrix and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\operatorname{ROC}(\operatorname{AUC}=0.9976)$</tex> analyses provide support for the reliability of the research findings. The framework developed is a reliable model to explicitly address the inherent data issues of imbalanced data and variability and provide support in improving early diagnosis of pneumonia in health care discrepancies and resource-poor settings. The ethical implications of this study, including privacy and fairness, support the responsible deployment of AI, and establishes the potential of its use in other areas of medicine and medical imaging.

Ähnliche Arbeiten