18F-FDG-PET/CT-based machine learning model evaluates indeterminate adrenal nodules in patients with extra-adrenal malignancies

Abstract

Background: To assess the value of an ¹⁸F-FDG-positron emission tomography/computed tomography (PET/CT)-based machine learning model for distinguishing between adrenal benign nodules (ABNs) and adrenal metastases (AMs) in patients with indeterminate adrenal nodules and extra-adrenal malignancies.

Methods: A total of 303 patients who underwent ¹⁸F-FDG-PET/CT with indeterminate adrenal nodules and extra-adrenal malignancies from March 2015 to June 2021 were included in this retrospective study (training dataset (n = 182): AMs (n = 97), ABNs (n = 85); testing dataset (n = 121): AMs (n = 68), ABNs (n = 55)). The clinical and PET/CT imaging features of the two groups were analyzed. The predictive model and simplified scoring system for distinguishing between AMs and ABNs were built based on clinical and PET/CT risk factors using multivariable logistic regression in the training cohort. The performances of the predictive model and simplified scoring system in both the training and testing cohorts were evaluated by the areas under the receiver operating characteristic curves (AUCs) and calibration curves. The comparison of AUCs was evaluated by the DeLong test.

Results: The predictive model included four risk factors: sex, the ratio of the maximum standardized uptake value (SUVmax) of adrenal lesions to the mean liver standardized uptake value, the value on unenhanced CT (CTU), and the clinical stage of extra-adrenal malignancies. The model achieved an AUC of 0.936 with a specificity, sensitivity and accuracy of 0.918, 0.835, and 0.874 in the training dataset, respectively, while it yielded an AUC of 0.931 with a specificity, sensitivity, and accuracy of 1.00, 0.735, and 0.851 in the testing dataset, respectively. The simplified scoring system had comparable diagnostic value to the predictive model in both the training (AUC 0.938, sensitivity: 0.825, specificity 0.953, accuracy 0.885; P = 0.5733) and testing (AUC 0.931, sensitivity 0.735, specificity 1.000, accuracy 0.851; P = 1.00) datasets.

Conclusions: Our study showed the potential ability of a machine learning model and a simplified scoring system based on clinical and 18F-FDG-PET/CT imaging features to predict AMs in patients with indeterminate adrenal nodules and extra-adrenal malignancies. The simplified scoring system is simple, convenient, and easy to popularize.

Keywords: Adrenal benign nodules; Adrenal metastases; Indeterminate adrenal nodules; Machine learning; PET/CT imaging.

Fig. 1

The process of dataset establishment, short time: within 6 months

Fig. 2

The overall workflow of the development and validation of the predictive model. First, the CT, PET, and clinical features were extracted from the training dataset, and then the predictive model was developed based on significant factors by logistic regression. Second, ROC and calibration curves were used to examine the performance of the nomogram both in the training and testing datasets. Third, a simplified scoring system was built based on the regression coefficients acquired from the training dataset for every individual feature in the predictive model and then the performance of this simplified scoring system was evaluated in both the training and testing datasets. Last, the simplified scoring tables were presented in both the training and testing datasets

Fig. 3

Strong correlations were observed among SUVmax, SUV/liver, and SUV/spleen

Fig. 4

A The AUC of the predictive model was 0.936 [95% CI 0.904–0.969], with a sensitivity, specificity and accuracy of 0.835, 0.918, and 0.874 in the training dataset, respectively. B The AUC of the predictive model was 0.931 [95% CI 0.889–0.973], with a sensitivity, specificity, and accuracy of 0.735, 1.000, and 0.851 in the testing dataset, respectively

Fig. 5

The AUC of the predictive model was higher than that of any feature alone

Fig. 6

Nomogram of the predictive model

Fig. 7

Good calibrations of the predictive model were shown in both the training (A) and testing datasets (B)

Fig. 8

The comparison of the AUCs between the predictive model and the simplified scoring system were not significantly different in the training dataset (A) and testing dataset (B)

Fig. 9

Good calibrations of the simplified scoring system were shown in both the training (A) and testing datasets (B)