Intelligent Diagnosis in Trauma: Exploring Machine Learning and Radiomics for Kidney Injury Assessment

Abstract

Purpose: The initial evaluation of trauma poses a formidable and time-intensive challenge. This study aims to scrutinize the diagnostic efficacy and utility of integrating machine learning models with radiomics features for the identification of blunt traumatic kidney injuries in abdominal CT images.

Materials and Methods: This investigation involved the collection of 600 CT scan images encompassing individuals with varying degrees of kidney damage resulting from trauma, as well as images from healthy subjects, sourced from the Kaggle dataset. An experienced radiologist performed the segmentation of axial images, and radiomics features were subsequently extracted from each region of interest. Initially, 30 machine learning models were deployed, with a final selection narrowed down to three models: Light Gradient-Boosting Machine (LGBM), Ridge Classifier, and Adaptive Boosting (AdaBoost). The performance of these chosen models was subjected to a more comprehensive examination.

Results: The AdaBoost model exhibited notable performance in diagnosing mild kidney injury, achieving accuracy and sensitivity rates of 93% and 94%, respectively. Furthermore, for severe kidney injury, the AdaBoost model demonstrated a remarkable sensitivity of 96% and an accuracy of 97%. The Area Under the Curve (AUC) values for this model were also calculated, yielding values of 92.91% and 97.04% for mild and severe renal injuries, respectively.

Conclusion: The artificial intelligence models employed in this study hold significant potential to enhance patient care by providing valuable assistance to radiologists and other medical professionals in the diagnosis and staging of trauma-related kidney injuries. These models offer the capability to prioritize positive studies, expedite evaluations, and accurately identify more severe injuries that may necessitate immediate intervention.