Competing at Teknofest Oncology was one of my first major experiences confronting the real potential of AI in medicine. The competition focused on detecting IDH (Isocitrate Dehydrogenase) mutation status from MRI scans of diffuse glioma patients. IDH mutation status is a critical biomarker directly affecting glioma prognosis and treatment selection — yet conventional methods require tissue biopsy. Our goal was to provide a non-invasive alternative.
The technical heart of the project was radiomics. Using the PyRadiomics library, I extracted over 100 features from MRI images: shape, texture, and intensity-based features. This stage turned out to be far more challenging than I anticipated — noise in images from different scanner protocols, missing data points, and non-normalized intensities were the main obstacles. In the preprocessing step I applied Z-score normalization, N4ITK bias correction and white matter normalization; I used HD-BET for skull stripping.
During model selection I compared Logistic Regression, Random Forest, SVM and XGBoost. To prevent overfitting on the small dataset, I addressed class imbalance with SMOTE and applied 5-fold cross-validation. The best AUC score was achieved with Random Forest (around 84%). Feature importance analysis revealed that textural features (GLCM, GLRLM) were significantly more informative than shape features — a finding that also aligns with the biological relationship between IDH mutation and tumor tissue heterogeneity.
The strongest impression the experience left on me was the ethical and technical responsibility of working with medical data. I experienced firsthand the limitations of a small dataset, class imbalance, and how much effort the long road to clinical validity demands. The Teknofest platform offered the opportunity to compare approaches from teams at different universities; examining the results of teams using transfer learning-based deep learning approaches in particular opened new questions about the complementarity of radiomics and DL methods.