MÜFİDE ÇİFTÇİ SÖZLÜ BİLDİRİ ÖDÜLLERİ

Yazol M.

34. Türk Nöroradyoloji Derneği Yıllık Toplantısı & 48. Esnr Yıllık Toplantısı & 15. Asya-Okyanusya Nöroradyoloji Ve Baş-Boyun Radyolojisi Kongresi, Eylül 2025

• Ödülün Kapsamı: Bilimsel/Mesleki Çalışmalardan Alınan Ödül
• Ödül Türü: Kongre, Konferans, Festival veya Sempozyum Kurullarınca Verilen Ödül
• Ödül Veren Ülke: Türkiye
• Ödülü Veren Organizasyon: 34. Türk Nöroradyoloji Derneği Yıllık Toplantısı & 48. Esnr Yıllık Toplantısı & 15. Asya-Okyanusya Nöroradyoloji Ve Baş-Boyun Radyolojisi Kongresi
• Araştırma Alanları: Tıp, Dahili Tıp Bilimleri, Radyodiagnostik, Sağlık Bilimleri
• Ödülün Tarihi: Eylül 2025
• Açıklama: <p>RADIOMICS-BASED CLASSIFICATION OF PEDIATRIC MEDULLOBLASTOMA AND POSTERIOR FOSSA EPENDYMOMA USING MRI</p><p>Introduction: Posterior fossa tumors (PFTs) represent approximately 50% of all pediatric brain tumors. Among these, medulloblastoma (MB) and ependymoma (EP) are the most common subtypes, frequently overlapping in both location and imaging characteristics—particularly within the fourth ventricle. Radiomics enables the extraction of quantitative features from MRI and holds promise for enhancing diagnostic differentiation. However, radiomics applications in pediatric PFTs remain limited. This study aims to identify radiomics-based differences between MB and EP and&nbsp;valuate the performance of machine learning techniques in their classification. Methods: The study included 26 patients (14 EP, 13 MB). Mean ages were 7.5 (EP) and 9.1 (MB). EPs were mostly in the fourth ventricle or pontocerebellar cistern; MBs were mainly in the fourth ventricle. MRI sequences (T1, T1C, T2, and FLAIR) were preprocessed using the CaPTk BraTS pipeline. Tumors were manually segmented into four subregions: cystic/necrotic core, peritumoral edema, non-enhancing tumpr and enhancing tumor. More than 100 radiomic features were extracted using PyRadiomics and subsequently normalized. Class imbalance was addressed using SMOTE and random undersampling (RUS). Feature selection was performed using random forest importance (RF_importance) and ReliefF methods. Three classifiers—Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP)—were evaluated using 5-fold crossvalidation and balanced accuracy as the performance metric. Results: In the nested cross-validation analysis, the highest sequence-level balanced accuracy was achieved using radiomic features from the T1 sequence (83%) with an SVM classifier. T1C and T2 sequences followed with 80% and 70% accuracy, respectively. At the subregion level, the best performance was observed in the necrotic region of the T1C (80%, SVM), followed by the enhancing tumoral region of T1C (77%, SVM), the non-enhancing tumoral region of T1 (73%, SVM), and the peritumoral edema of FLAIR (70%, RF). These findings suggest that certain tumor subregions provide distinct radiomic signatures for differentiating EPs from MBs. The lowest performance was seen in the non-enhancing region of T2 (47%, SVM). Discussion and Conclusion: This study demonstrates that MRI-based radiomic features, combined with machine learning, can differentiate between pediatric medulloblastomas and ependymomas by capturing tumor heterogeneity. Among tumor subregions, necrotic and enhancing components demonstrated the highest discriminative performance, while non-enhancing regions yielded lower accuracy, possibly reflecting greater internal heterogeneity. Limitations include small sample size and lack of molecular data. Future multicenter studies with molecular profiling may improve diagnostic accuracy and biological interpretation.</p>