RESUMO
OBJECTIVES: In this single-center study, we aimed to propose a machine-learning model and assess its ability with clinical data to classify low- and high-risk thymoma on fluorine-18 (18F) fluorodeoxyglucose (FDG) (18F-FDG) PET/computed tomography (CT) images. METHODS: Twenty-seven patients (14 male, 13 female; mean age: 49.6 ± 10.2 years) who underwent PET/CT to evaluate the suspected anterior mediastinal mass and histopathologically diagnosed with thymoma were included. On 18F-FDG PET/CT images, the anterior mediastinal tumor was segmented. Standardized uptake value (SUV)max, SUVmean, SUVpeak, MTV and total lesion glycolysis of primary mediastinal lesions were calculated. For texture analysis first, second, and higher-order texture features were calculated. Clinical information includes gender, age, myasthenia gravis status; serum levels of lactate dehydrogenase (LDH), alkaline phosphatase, C-reactive protein, hemoglobin, white blood cell, lymphocyte and platelet counts were included in the analysis. RESULTS: Histopathologic examination was consistent with low risk and high-risk thymoma in 15 cases and 12 cases, respectively. The age and myasthenic syndrome were statistically significant in both groups (P = 0.039 and P = 0.05, respectively). The serum LDH level was also statistically significant in both groups (450.86 ± 487.07 vs. 204.82 ± 59.04; P < 0.001). The highest AUC has been achieved with MLP Classifier (ANN) machine learning method, with a range of 0.830 then the other learning classifiers. Three features were identified to differentiate low- and high-risk thymoma for the machine learning, namely; myasthenia gravis, LDH, SHAPE_Sphericity [only for 3D ROI (nz>1)]. CONCLUSIONS: This small dataset study has proposed a machine-learning model by MLP Classifier (ANN) analysis on 18F-FDG PET/CT images, which can predict low risk and high-risk thymoma. This study also demonstrated that the combination of clinical data and specific PET/CT-based radiomic features with image variables can predict thymoma risk groups. However, these results should be supported by studies with larger dataset.