Automatic Multiphase CT Classification Using Machine Learning Methods

Sammanfattning

Computed Tomography (CT) is widely used in clinical diagnostics to visualise internal structures and detect abnormalities in the body. In contrast-enhanced CT (CECT), a contrast medium is injected into the bloodstream to enhance specific organs and tissues. Since different anatomical structures are enhanced at different time points, CECT images are commonly acquired in specific contrast phases. However, variations in patient-specific contrast circulation make it challenging to consistently capture the intended phase, potentially affecting diagnostic quality. This study investigates a machine learning based approach for automatic CT phase classification to support verification of acquired images. A stepwise pipeline was developed. First, anatomical structures such as the aorta, kidneys, liver, spleen, urinary bladder, and portal vein were segmented using an open source deep learning model called TotalSegmentator. Radiomic feature descriptors were then extracted from the segmented regions. These features were used to train and compare multiple machine learning models, including Logistic Regression, Random Forest, Support Vector Machine, and Extreme Gradient Boosting. Among the evaluated models, Extreme Gradient Boosting achieved the best performance, with a macro F1-score of 0.79 on the external test set and 0.90 on an additional dataset from Karolinska University Hospital. The results varied between the datasets, indicating that model performance depends on the characteristics of the input data. Feature importance was analysed using SHapley Additive exPlanations (SHAP). The most important features were associated with organ-specific enhancement patterns relevant to each phase and were consistent with radiological expectations. These included the absence of contrast enhancement in the non-contrast phase, aortic enhancement in the arterial phase, portal vein- and liver-related features in the portal venous phase, and kidney-related features in the delayed phase. The results show that automatic CT phase classification using organ-based features is possible. However, further validation on larger and more diverse datasets is needed before clinical use. A complete pipeline implementing the method was also developed and made available for external testing in the following GitHub repository: https://github.com/norael-zein/MainPhaseClassifier