OS9.6 Diagnosis of pseudoprogression using FET PET radiomics

• Published in: Neuro-oncology (Charlottesville, Va.) Va.), 2019-09, Vol.21 (Supplement_3), p.iii19-iii19
• Main Authors: Lohmann, P, Elahmadawy, M A, Werner, J, Rapp, M, Ceccon, G, Fink, G R, Shah, N J, Langen, K, Galldiks, N
• Format: Article
• Language: English
• Subjects: Oral Presentations
• ISSN: 1522-8517; 1523-5866
• DOI: 10.1093/neuonc/noz126.064

Abstract BACKGROUND Radiomics derived from different imaging modalities is gaining increasing interest in the field of neuro-oncology. Besides MRI, amino acid PET radiomics may also improve the to date challenging, clinically relevant diagnostic problem of differentiating pseudoprogression (PsP) from tumor progression (TP). To this end, we here explored the potential of O-(2-[18F]fluoroethyl)-L-tyrosine (FET) PET radiomics to discriminate between PsP and TP. MATERIAL AND METHODS Thirty-five newly diagnosed IDH-wildtype glioblastoma patients with MRI findings suspicious for TP within 12 weeks after completion of chemoradiation with temozolomide underwent an additional dynamic FET PET scan. FET PET tumor volumes were segmented using a tumor-to-brain ratio (TBR) ≥ 1.6. The static PET parameters TBRmax and TBRmean, as well as the dynamic parameter time-to-peak (TTP), were calculated. For radiomics analysis, the number of datasets for model generation was increased using data augmentation techniques. Subsequently, 70 datasets were available for model generation. Prior to further processing, patients were randomly assigned to a discovery and a validation dataset in a ratio of 70/30, with balanced distribution of PsP and TP diagnoses. Forty-two radiomics features (4 shape-based, 6 first- and 32 second-order features) were obtained using the software LifeX (lifexsoft.org). Afterwards, a z-score transformation was performed for data normalization. For feature selection, recursive feature elimination using random forest regressors was performed. For the final model generation, the number of parameters was limited to three to avoid data overfitting. Different algorithms for model calculation were compared, and the diagnostic accuracy was assessed using leave-one-out cross-validation. Finally, the resulting models were applied to the validation dataset to evaluate model robustness. RESULTS Eighteen patients were diagnosed with TP, and 17 patients had PsP. Diagnoses were based on a neuropathological confirmation or clinicoradiological follow-up (26% and 74%, respectively). The diagnostic accuracy of the best single FET PET parameter was 75% (TBRmax). Combining TBRmax and TTP increased the diagnostic accuracy to 83%. Other combinations of static and dynamic FET PET parameters, however, did not further increase the accuracy. The highest diagnostic accuracy of 92% was achieved by a three-parameter model combining the FET PET parameter TTP with two radiomics features. The