Predictive AI modeling of brain tumor MRI based on hybrid nonlinear fuzzy regression

The MRI imaging of brain tumours cannot be properly and reliably diagnosed because of high intra-class variability, imaging noise, and limited interpretability of deep learning models. Even though convolutional neural networks (CNNs) have attained great predictive accuracy, their black-box and absence of uncertainty measures limit clinical trust and adoption. This study proposes a Non-linear Hybrid Fuzzy Regression Model (NHFRM) that integrates deep CNN-based feature extraction with nonlinear fuzzy regression to collaboratively focus on prediction accuracy, explainability and model uncertainty. The suggested scheme embodies the regression coefficients expressed in terms of more than one fuzzy paradigm, such as triangular, trapezoidal, intuitionistic, and Pythagorean fuzzy numbers, which allows clear bands of uncertainty around the predictions. Non-linear interaction between features is captured by kernel regularization, whereas temporal smoothness formulation is proposed for future longitudinal MRI application. The experiments performed on a publicly available brain tumor MRI dataset of 7,023 images show that NHFRM outperformed traditional regression approaches, ensemble techniques, and deep learning models in terms of accuracy of prediction, discrimination (AUC = 0.92) and calibration. The importance of regularization of kernels and penalties on the spread of kernels have been shown to aid in enhancing robustness in a study of ablation. The proposed model provides clinical readers with a meaningful decision support tool based on calibrated predictions, uncertainty intervals defined, and is therefore able to offer a solution to bridging the assumptions of black-box deep learning and explainable medical artificial intelligence.