Two-Stage CT Subphase Classification Using Organ-Level Features

Sammanfattning

Accurate identification of contrast-enhanced computed tomography (CT) phases is important for ensuring that acquired images are suitable for their intended diagnostic purpose. While automated methods have previously been applied to broader CT phase classification, distinguishing clinically relevant subphases remains challenging due to subtle differences in contrast enhancement and variation between scanners, protocols, and patient populations. This thesis developed and evaluated machine learning models for automatic classification of CT contrast phases and subphases using organ-level quantitative features extracted from segmented anatomical structures. A hierarchical classification pipeline was implemented, consisting of a main model for the non-contrast (NP), arterial phase (AP), portal-venous phase (VP), and delayed phase (DP), followed by submodels for early and late arterial phase (EAP and LAP) and nephrographic (NEP) phase. Several machine learning algorithms were compared, including Logistic Regression, Support Vector Machine, Random Forest, and XGBoost. The final models were evaluated on an internal test set and on two external datasets, including VinDr Multiphase and a dataset from Karolinska University Hospital. The main classification model achieved strong performance, with macro F1-scores of 0.88 on the internal test set, 0.95 on the external VinDr dataset, and 0.93 on the dataset from Karolinska University Hospital. Performance was lower for the subphase models. The EAP versus LAP model achieved macro F1-scores of 0.81 on the internal test set and 0.76 on the external VinDr dataset. A complete evaluation was not possible on the Karolinska University Hospital dataset because no ground-truth LAP cases were available. However, the model achieved an F1-score of 0.94 for the 18 available EAP cases. The NEP versus VP model achieved macro F1-scores of 0.81, 0.74, and 0.68 on the same datasets, respectively. Tree-based ensemble models, particularly XGBoost and Random Forest, generally outperformed Logistic Regression and Support Vector Machine. Model interpretation using SHAP indicated that the models relied on physiologically relevant organ-level features. The results show that organ-level quantitative features can support automatic CT phase classification, especially for broader contrast phases. However, distinguishing neighboring subphases remains more challenging and was less robust across external datasets. The proposed pipeline may support automatic phase verification, but further validation on larger and more diverse clinical datasets is needed. The constructed pipeline is available in the following Github repository: https://github.com/norael-zein/SubPhaseClassifier