Machine learning for differentiation of lipid-poor adrenal adenoma and subclinical pheochromocytoma based on multiphase CT imaging radiomics

Abstract

Abstract Objective The aim of this study was to use radiomics analysis of multiphase computed tomography (CT) imaging to develop and validate machine learning models that can accurately differentiate between lipid-poor adrenal adenoma (LPA) and subclinical pheochromocytoma (sPHEO) to improve the accuracy of preoperative diagnosis of the two.Methods A retrospective analysis was performed on 134 patients who underwent abdominal multiphase spiral CT scans in three local tertiary hospitals between March 2015 and November 2022. The analysis included 74 cases of LPA (52 at our hospital and 22 at other hospitals) and 60 cases of sPHEO (44 at our hospital and 16 at other hospitals), all of which were surgically and pathologically confirmed. Tumors were delineated using 3D Slicer and radiomics were extracted using PyRadiomics, a plug-in to the software. Cases from internal hospital were randomly split into training and test sets in a 7:3 ratio, while all cases from external hospitals were used as the validation set. The T-test and the minimum absolute contraction and selection operator (LASSO) were used to reduce the dimensionality of the data. Then six dichotomous models were developed, including k-nearest neighbor (KNN), logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), and multi-layer perceptron (MLP). The diagnostic performance of each model was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC). The ROC curves of the test and validation sets were compared using DeLong's method to determine the most effective model for distinguishing between LPA and sPHEO.Results A total of 902 rows and 112 columns of radiomic feature data were extracted from multiple phases and slice-thickness CT data. After dimensionality reduction processing, 13 - dimensional radiomic feature data was obtained. The six binary models demonstrated good diagnostic performance for each phase and slice thickness, as well as for the entire CT data, with AUC values ranging from 0.706 to 1. Among these models, RF, SVM, and MLP showed particularly good diagnostic performance. The ROC curves of RF, SVM, and MLP did not show a statistically significant difference (p < 0.05) for different phase, slice-thicknesses, as well as the entire test and validation sets, except for the thick slice-thickness data sets. The AUC value of the MLP model for the non-contrast CT validation set was 0.979, which is quite high. Furthermore, there was no significant difference in the ROC curves when compared to other phases and the entire validation sets (p < 0.05).Conclusions The CT radiomics-based machine learning model was able to differentiate between LPA and sPHEO well, even using non-contrast CT data alone to efficiently discriminate between the two.