Explainable deep learning for rib fracture detection in chest x-rays

Detecting rib fractures is a crucial and specialized task in the medical field due to the complications that can arise from untreated fractures, such as symptom relief challenges, preventing further issues, increased healthcare costs, and overall patient health. Deep learning models have been developed to enhance accurate rib fracture detection, provide support towards helping radiologist efficiency, and provide as an option in situations where radiologists are not immediately available. However, the reliability of these models has often been questioned, as previous studies have reported only overall accuracy without assessing the model's focus on the correct regions, thus lacking explainability. In this study, we aim to improve overall accuracy and elucidate the model's focus when detecting rib fractures. We utilized GRAD-CAM to visualize the regions of interest identified by the model. We used the DenseNet121 architecture, performed data preprocessing, including image cropping and data augmentation, and conducted extensive experimentation and hyperparameter tuning. Our optimized model achieved 74.9% accuracy on the CheXpert testing dataset, 84.5% validation accuracy on the Medical College of Wisconsin (MCW) dataset, and an AUROC score of 0.82. Grad-CAM visualization revealed issues with the model's focus and localization abilities, highlighting current limitations and providing areas for future research. Our work demonstrates the potential of deep learning for automated rib fracture detection in chest X-rays and underscores the need for further improvements in model explainability.