Glioma Grade Classification: Hypertuned CNN With Saliency-Based Insight From MRI

Glioma classification is critical for early diagnosis and treatment planning, yet manual MRI-based diagnosis is time-consuming and prone to errors. This study proposed a CNN-based approach optimized with KerasTuner for glioma classification, focusing on achieving high accuracy using a lightweight architecture. While hyperparameter tuning is well-established, its tailored application to this dataset and problem domain improves performance and reproducibility. The model employs KerasTuner to optimize key hyperparameters, including filter sizes, pooling layers, dropout rates, and optimizers, ensuring improved feature extraction and robustness. A dataset comprising 16,956 T2w Flair MRI slice images from 231 patients was used and categorized into three tumor grades (Grade 2, Grade 3, and Grade 4). The proposed model was trained using five-fold cross-validation, achieving an average multi-class classification accuracy of 89% and a binary classification accuracy of 93% in distinguishing low-grade (Grades 2) from high-grade (Grade 3 and Grade 4) tumors. Comparative analysis with state-of-the-art deep learning models such as ResNet-50, DenseNet-201, and EfficientNet-B0 demonstrated that our optimized CNN outperformed these architectures in accuracy, recall, and F1-score. Moreover, an ablation study and statistical analysis were performed. The experimental results confirm that integrating hyperparameter tuning in lightweight CNN models significantly improves tumor classification performance, reducing misclassification errors between adjacent grades. These findings emphasize the potential of deep learning in automated glioma classification, providing a more efficient and reliable tool for clinical diagnosis. Saliency map visualizations revealed that the model predominantly attends to the peritumoral regions, highlighting their significance in accurate glioma grade classification. Our hyper-tuned CNN achieved comparable performance across T1-weighted, T1-weighted-Gd, T2-weighted, and FLAIR inputs. Extensive augmentation trials, dropout/depth ablation, and paired-t analysis (t = 4.37, p < 0.0001) verified the model’s superiority over other baselines. Future work will explore multimodal imaging, domain adaptation, and transfer learning to enhance model robustness across diverse medical datasets.