Abstract
Abstract
Glioma is one of the most prevalent primary tumours of the central nervous system, and its Isocitrate Dehydrogenase(IDH) molecular subtype is directly linked to patient prognosis, survival, and treatment plans. However, there are currently no reliable, non-invasive techniques for preoperative prediction, and postoperative pathology and genetic testing are the only ways to identify IDH molecular subtypes.Radiomic characteristics are very useful for characterising tumour heterogeneity and can quantitatively extract high-dimensional information from MRI images, such as texture, intensity, and shape. In evaluating glioma aggressiveness, growth patterns, and angiogenesis, spatial semantic data successfully supplement radiomic features and provide improved clinical interpretability and stability.In order to accurately predict IDH molecular subtypes in gliomas, this work suggests a collaborative analysis approach that combines multi-parameter MRI radiomic features, spatial semantic features, and clinical data. This technique builds spatial semantic features by combining morphological analysis, gradient statistics, and 3D geometric modelling, and extracts multi-parameter MRI radiomic characteristics utilising a multi-scale texture and morphological feature extraction strategy based on the IBSI standard. To identify a stable and optimal subset of features, we employed a multi-stage feature selection method. In order to accurately predict the IDH molecular subtypes of gliomas, random forests were finally used to combine clinical data with imaging attributes.According to the experimental findings, the suggested model's best AUC in five-fold cross-validation was 96.12%. The model's final test results showed an AUC of 95.62%, an ACC of 91.67%, a sensitivity of 88.89%, and a specificity of 93.94%.