A Classifier Model Based on CT Data from Different CTPhases for Distinguishing LPAs and PCCs

Background: Lipid-poor adrenal adenomas (LPAs) and pheochromocytomas (PCCs) are similar tumours, but misdiagnosed LPAs may lead to health risks such as hypertensive crisis due to improper treatment. The aim of this study was to develop an efficient method for classifying LPAs and PCCs on the basis of different CT scans that minimises the number of radiation doses. Methods: The patients included in this study were randomly divided into training and validation groups (the ratio was 7:3). The datasets, including 2-(plain and venous enhanced CT scans) or 3-phase CT data, were separately used to construct XGBoost, Gradient Boosted Decision Tree (GBDT), AdaBoost, random forest and decision-tree models. Receiver operator characteristic (ROC) curves were used to evaluate the models, and the DeLong test was used to determine significant differences. Results: The models constructed were XGBoost, GBDT, AdaBoost, random forest and decision tree and their efficacies Area Under the Curves (AUCs) in the 2-phase CT group were 0.91, 0.89, 0.85, 0.78, and 0.71, respectively, while those in the 3-phase CT group were 0.92, 0.91, 0.89, 0.81, and 0.78, respectively. The optimal model in both the 2-and 3-phase groups was XGBoost; this model exhibited similar performance in both groups. The DeLong test also confirmed some difference in XGBoost between the two groups. Conclusions: Our XGBoost-based model constructed using 2-phase CT data is similar to that constructed using 3-phase CT data; both of them exhibited good performance in the classification of LPAs and PCCs.