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

Area of Science:

• Radiology and Oncologic Imaging
• Machine Learning in Healthcare
• Medical Diagnostics

Background:

• Lipid-poor adrenal adenomas (LPAs) and pheochromocytomas (PCCs) are distinct adrenal tumors with overlapping imaging characteristics.
• Misdiagnosis can lead to severe health risks, including hypertensive crisis from inappropriate treatment of LPAs.
• Accurate differentiation is crucial for patient management and treatment planning.

Purpose of the Study:

• To develop and evaluate an efficient machine learning model for classifying LPAs and PCCs.
• To compare the diagnostic performance of models using different phases of CT scans.
• To minimize radiation dose by exploring the efficacy of 2-phase versus 3-phase CT protocols.

Main Methods:

• Patients were randomly assigned to training (70%) and validation (30%) groups.
• XGBoost, GBDT, AdaBoost, random forest, and decision tree models were trained using 2-phase (plain and venous enhanced) and 3-phase CT data.
• Model performance was assessed using Receiver Operator Characteristic (ROC) curves and the DeLong test for significance.

Main Results:

• XGBoost demonstrated the highest efficacy in both 2-phase (AUC=0.91) and 3-phase (AUC=0.92) CT groups.
• The XGBoost model showed comparable performance across both 2-phase and 3-phase CT datasets.
• Other models like GBDT, AdaBoost, random forest, and decision tree showed lower AUC values in both protocols.

Conclusions:

• An XGBoost-based classification model using 2-phase CT data achieves performance comparable to 3-phase CT.
• This approach offers an efficient and potentially lower-radiation method for differentiating LPAs and PCCs.
• The findings support the use of XGBoost with reduced CT phases for accurate adrenal tumor classification.