A Self Explainable Graph EfficientNet Framework and Deep Survival Modeling for Brain Tumor classification and Prognosis

Abstract

Clinical neuro-oncology faces difficulties classifying and predicting prognosis of brain tumors. These difficulties stem from heterogeneous tumor morphology, like image data noise and limited deep learning model interpretability. The study proposes a novel framework for accurately classifying brain tumors as well as predicting prognosis named the Self-Explainable Graph EfficientNet model (SEGE-Net) that incorporates a deep survival model into its architecture. First adaptive Kalman filtering (AKF) is applied to multivariate Magnetic Resonance Imaging (MRI) scans to reduce noise and stabilize image intensity enhancing the quality of MRI scans. Second a modified version of the context-aware feature pyramid network (CFPNet-M) is used to segment the tumors sub regions in order to extract the most discriminative features and improve generalization through geometric augmentation. Next a Global Binary Pattern (GBP) and Completed Local Binary Pattern (CLBP) method is used for texture feature extraction. SEGE-Net combines the EfficientNet model with self-explainable graph-based neural networks (GNN) to model the relationships between extracted textures using discriminative radiomic features. Finally DeepSurv was used to generate patient specific survival probabilities and risk scores. The proposed method achieved segmentation score for precision is 0.93, recall is 0.90 and classification for accuracy is 98.2% and F1-score is 99%.Through experimentation on the BRATS 2020 dataset and Brain Tumor MRI dataset the SEGE-Net provided improved classification accuracy, reliable survival estimates and enhanced explainability of the models prediction processes