OPTIMAL BINARIZATION METHOD FOR IBSI FEATURES IN INTERPRETABLE BRAIN TUMOR DIAGNOSIS

Abstract

This paper presents an optimal binarization method for standardized IBSI features in creating interpretable diagnostic models for brain tumors. The growing adoption of radiomics in medical imaging has generated extensive quantitative features describing texture, morphology, and intensity properties of pathological formations. However, integration of these continuous numerical characteristics into interpretable rule-based diagnostic systems requires effective discretization approaches that preserve diagnostic information while maintaining clinical interpretability. The proposed approach is based on mutual information maximization criterion considering local diagnostic context and ensures optimal balance between feature informativeness and compatibility with logical rule architecture. An adaptive threshold determination procedure has been developed through discrete search over candidate set with composite criterion application that accounts for both information value and distribution balance of binarized values. The methodology incorporates percentile-based thresholds and statistically grounded values to form a comprehensive candidate set, enabling robust feature transformation across diverse clinical scenarios. Experimental validation on MRI dataset with four pathology classes (glioma, meningioma, pituitary tumor, and no tumor) containing 64 IBSI-standardized features showed that optimized thresholds provide average mutual information of 0.342 bits compared to 0.287 bits for fixed median thresholds, representing 19.2% improvement with statistical significance (p < 0.001). The proposed method ensures generation of more stable and clinically relevant diagnostic rules through preservation of medical meaningfulness of selected features during their transformation to binary format. Integration with Decision Rules Network architecture demonstrated 7.3% accuracy improvement and achieved 89.3% local consistency with base VGG-16 model. The research addresses fundamental challenge of feature discretization for rule-based interpretable systems in medical imaging by developing theoretically grounded optimization framework that maintains both mathematical rigor and clinical applicability of resulting binary features, facilitating transparent and trustworthy AI-assisted diagnostic decision-making in clinical practice.