Scholarlink

<p>Brain tumors pose a major global health challenge, requiring accurate and early detection to support effective treatment and improve patient survival. This study evaluates and compares the performance of various machine learning and deep learning models for brain tumor classification using magnetic resonance imaging (MRI) data from two independent datasets (Dataset A and Dataset B). The objective is to identify the most efficient approach for enhancing diagnostic accuracy and model generalization. Nine models are developed and assessed: a Convolutional Neural Network (CNN) trained from scratch, a hybrid CNN–SVM model, Xception, Vision Transformer (ViT), Support Vector Machine (SVM), Voting Classifier, Bagging Classifier, AdaBoost Classifier, and Gradient Boosting Classifier. To enhance model interpretability, four explainable artificial intelligence (XAI) techniques: Integrated Gradients (IG), Smooth Integrated Gradients (Smooth IG), Guided Integrated Gradients (Guided IG), and Superimposed Integrated Gradients (Superimposed IG), are employed. All models undergo a standardized preprocessing pipeline including image resizing, rescaling, dataset splitting, and oversampling to address class imbalance. Furthermore, Principal Component Analysis (PCA) is applied prior to training the SVM, Voting, Bagging, AdaBoost, and Gradient Boosting classifiers to reduce dimensionality and improve computational efficiency. Experimental results demonstrate that the Xception model, leveraging transfer learning, achieves 97% accuracy, 97% precision, 97% recall, and a 97% F1 score on Dataset A, while the CNN trained from scratch achieves 99% accuracy, 98% precision, 99% recall, and a 98% F1 score on Dataset B. These results confirm that the proposed architectures surpass existing state-of-the-art methods, providing a robust and interpretable framework for automated brain tumor classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable MRI-based brain tumor classification using deep learning and hybrid models with integrated gradients

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Samar Farhani,
  • Amira Echtioui

摘要

Brain tumors pose a major global health challenge, requiring accurate and early detection to support effective treatment and improve patient survival. This study evaluates and compares the performance of various machine learning and deep learning models for brain tumor classification using magnetic resonance imaging (MRI) data from two independent datasets (Dataset A and Dataset B). The objective is to identify the most efficient approach for enhancing diagnostic accuracy and model generalization. Nine models are developed and assessed: a Convolutional Neural Network (CNN) trained from scratch, a hybrid CNN–SVM model, Xception, Vision Transformer (ViT), Support Vector Machine (SVM), Voting Classifier, Bagging Classifier, AdaBoost Classifier, and Gradient Boosting Classifier. To enhance model interpretability, four explainable artificial intelligence (XAI) techniques: Integrated Gradients (IG), Smooth Integrated Gradients (Smooth IG), Guided Integrated Gradients (Guided IG), and Superimposed Integrated Gradients (Superimposed IG), are employed. All models undergo a standardized preprocessing pipeline including image resizing, rescaling, dataset splitting, and oversampling to address class imbalance. Furthermore, Principal Component Analysis (PCA) is applied prior to training the SVM, Voting, Bagging, AdaBoost, and Gradient Boosting classifiers to reduce dimensionality and improve computational efficiency. Experimental results demonstrate that the Xception model, leveraging transfer learning, achieves 97% accuracy, 97% precision, 97% recall, and a 97% F1 score on Dataset A, while the CNN trained from scratch achieves 99% accuracy, 98% precision, 99% recall, and a 98% F1 score on Dataset B. These results confirm that the proposed architectures surpass existing state-of-the-art methods, providing a robust and interpretable framework for automated brain tumor classification.