Static FET PET radiomics for the differentiation of treatment-related changes from glioma progression

Abstract

Purpose: To investigate the potential of radiomics applied to static clinical PET data using the tracer O-(2-[¹⁸F]fluoroethyl)-L-tyrosine (FET) to differentiate treatment-related changes (TRC) from tumor progression (TP) in patients with gliomas.

Patients and methods: One hundred fifty-one (151) patients with histologically confirmed gliomas and post-therapeutic progressive MRI findings according to the response assessment in neuro-oncology criteria underwent a dynamic amino acid PET scan using the tracer O-(2-[¹⁸F]fluoroethyl)-L-tyrosine (FET). Thereof, 124 patients were investigated on a stand-alone PET scanner (data used for model development and validation), and 27 patients on a hybrid PET/MRI scanner (data used for model testing). Mean and maximum tumor to brain ratios (TBR_mean, TBR_max) were calculated using the PET data from 20 to 40 min after tracer injection. Logistic regression models were evaluated for the FET PET parameters TBR_mean, TBR_max, and for radiomics features of the tumor areas as well as combinations thereof to differentiate between TP and TRC. The best performing models in the validation dataset were finally applied to the test dataset. The diagnostic performance was assessed by receiver operating characteristic analysis.

Results: Thirty-seven patients (25%) were diagnosed with TRC, and 114 (75%) with TP. The logistic regression model comprising the conventional FET PET parameters TBR_mean and TBR_max resulted in an AUC of 0.78 in both the validation (sensitivity, 64%; specificity, 80%) and the test dataset (sensitivity, 64%; specificity, 80%). The model combining the conventional FET PET parameters and two radiomics features yielded the best diagnostic performance in the validation dataset (AUC, 0.92; sensitivity, 91%; specificity, 80%) and demonstrated its generalizability in the independent test dataset (AUC, 0.85; sensitivity, 81%; specificity, 70%).

Conclusion: The developed radiomics classifier allows the differentiation between TRC and TP in pretreated gliomas based on routinely acquired static FET PET scans with a high diagnostic accuracy.

Keywords: Amino acid PET; Artificial intelligence (AI); Brain tumors; Machine learning.