Enhanced Diagnostic of Pulmonary Embolism Detection using DenseNet and XGBoost

Abstract

Pulmonary embolism (PE) poses a significant medical challenge, often linked to severe morbidity and mortality rates. This study investigates the effectiveness of machine learning techniques in detecting pulmonary embolism through radiological images. Utilizing DenseNet121 convolutional neural networks and XGBoost classifiers, we analyze the RSNA Kaggle Pulmonary Embolism Dataset, encompassing diverse CT scans. Initial preprocessing involves standardizing image sizes and normalization, with a focus on CT scan image origins and Hounsfield Units (HU) augmentation. Features extracted using DenseNet121 are utilized for classification tasks via an XGBoost classifier. Comprehensive evaluation metrics including ROC curves, AUC scores, confusion matrices, and classification reports assess model performance. Results show promising mean accuracy rates (99.38%), sensitivity (99.40%), and specificity (99.40%), highlighting the potential of machine learning in assisting clinicians with precise PE predictions from thoracic CT scans. This research emphasizes the practicality of machine learning in enhancing PE diagnosis through radiological imaging, contributing to improved diagnostic precision and timely clinical interventions.