Enhancing Glioma Grading with Texture Features and SIHO-Enhanced Gradient Boosting

Abstract

Accurate grading of gliomas is critical for predicting tumor behavior and informing treatment decisions. However, traditional approaches may poorly characterize tumor hetero geneity and generally lack interpretability. In this study, we proposed an automated glioma grading framework combining de noising, hybrid segmentation, radiomic feature extraction, and LightGBM classification optimized using the Spuggy Interactive Hyperparameter Optimization (SIHO) algorithm. The novelty of SIHO is its interactive hyperparameter search mechanism, which balances exploration and exploitation to select compact and effective LightGBM hyperparameters for improved glioma grading. We built our framework on the BraTS 2019 dataset which contains a total of 335 patient cases with 259 cases of high grade glioma (HGG) and 76 cases of low-grade glioma (LGG) with T1, T1ce, T2 and FLAIR MRI modalities, as well as ground truth masks. To improve validation rigor, both an 80:20 train test split and k-fold cross-validation were used. Class imbalance is handled by stratified sampling and performance was assessed using accuracy, sensitivity, and specificity instead of accuracy alone. In the 80:20 evaluation, the proposed SIHO–LightGBM model achieved 96.59% accuracy, 94.73% sensitivity, and 96.98% specificity. The experimental results reveal the potential of our proposed framework for automated glioma grading.

Keywords