Radiomics-based machine learning for splenic injury diagnosis using computed tomography (CT) images.
Authors
Affiliations (3)
Affiliations (3)
- Student Research Committee, Kermanshah University of Medical Sciences, Kermanshah, Iran.
- Department of Medical Physics, School of Medicine, Kermanshah University of Medical Sciences, Kermanshah, Iran. [email protected].
- Modeling in Health Research Center, Institute for Futures Studies in Health, Kerman University of Medical Sciences, Kerman, Iran.
Abstract
Efficient trauma assessment is essential for optimal patient care, with imaging playing a critical role in the detection of injuries. Rapid and accurate classification of traumatic spleen injuries is critical for clinical decision-making; however, manual assessment of CT images can be subjective and time-consuming, highlighting the need for objective and automated diagnostic tools. This study aims to evaluate the impact of machine learning models and radiomics features in diagnosing splenic trauma lesions on computed tomography images. A dataset of 600 computed tomography images, including individuals with mild and severe traumatic spleen injuries as well as healthy controls-was collected from the Kaggle database. An experienced radiologist segmented the axial images, and radiomics features were extracted from each designated region of interest for further analysis. Initially, 25 machine learning models were evaluated; ultimately, three-Light Gradient Boosting Machine, Ridge Classifier, and Adaptive Boosting-were selected for detailed assessment. Model performance was measured using accuracy, precision, sensitivity, specificity, area under the receiver operating characteristic curve, F1 score, and misclassification rate. The Light Gradient Boosting Machine model exhibited superior effectiveness in diagnosing mild spleen injuries, achieving an accuracy of 98%, precision, and specificity of 100%. Meanwhile, the Adaptive Boosting model demonstrated acceptable performance in diagnosing severe injuries, achieving an accuracy of 90%, precision of 92.15%, and specificity of 91%. These machine learning models exhibited remarkable capability in automatically detecting traumatic spleen injuries on abdominal computed tomography scans. By integrating radiologist expertise into the analytical framework, our method enables rapid pre-screening of a large number of cases for spleen lesions.