OS6.3 Radiomics and machine learning on [(18)F]FET PET and T1ce MRI discriminate between low-grade and high-grade glioma

BACKGROUND: Gliomas are the most frequent malignant primary brain tumours in adults, accounting for about 70% of adult malignant primary brain tumours. They can be categorised into low-grade gliomas (LGG, WHO grade I-II) or high-grade gliomas (HGG, WHO grade III-IV). Accurate and early diagnosis of these tumours, for which histopathological analysis is the gold standard, is key to determine the optimal treatment procedure. Because biopsy and tumour resection are invasive procedures, there has been a growing interest in automated classification of brain tumours based on medical imaging. This retrospective study intends to evaluate automatic tumour segmentation and radiomics features from [(18)F]FET PET and contrast-enhanced T1-weighted (T1ce) MRI images for discrimination between LGG and HGG. MATERIAL AND METHODS: A pretreatment [(18)F]FET PET scan and anatomical MRI were performed in 30 patients. Histopathological analysis led to the diagnosis of 14 LGG and 16 HGG patients. An in-house developed and previously validated segmentation algorithm was used to automatically delineate the tumour simultaneously on T1ce and FLAIR MRI. This resulted in the identification of four tumour-related tissues: necrosis, oedema, non-enhancing tumour and contrast-enhancing tumour. These different tumour masks were used to extract 2913 quantitative radiomics features per patient from the T1ce and PET images. For each feature, a two sample t-test was performed between LGG and HGG patients as an initial feature selection. The 150 features with the lowest p-values were retained. Next, forward sequential selection and five-fold cross-validation was implemented to identify up to 15 features that are able to make the optimal distinction between LGG and HGG using a Random Forests classification model. The accuracy was assessed using leave-one-out validation. This procedure was performed three times: using only T1ce features, only PET features, and both T1ce and [(18)F]FET features. RESULTS: We achieve accuracies of 86.6%, 83.3% and 96.7% for the classification tasks using MRI-based, PET-based and combined features, respectively. The best result is obtained when combining one texture feature calculated on T1ce in the oedema and non-enhancing tumour region with two histogram and five texture features calculated on [(18)F]FET PET, both in the oedematous and total abnormal regions. This model achieves an excellent discrimination between LGG and HGG (accuracy 96.7%, AUC=0.951), using a minimal amount of features. We further observe that many predictive features are based on infiltrating rather than tumour core regions. CONCLUSION: This study illustrates that automatic tumour segmentation and extraction of radiomics features from combined [(18)F]FET PET and T1ce MRI scans are able to discriminate between LGG and HGG. Further evaluation on larger datasets is however necessary to validate this result.