Scholarlink

<p>The identification and categorization of brain tumors from magnetic resonance imaging (MRI) scans pose a significant challenge in neuro-oncology, where prompt and precise diagnosis is crucial for enhancing patient survival and informing treatment approaches. Recent advancements in deep learning have markedly improved the precision and efficacy of automated tumor diagnosis; nonetheless, several cutting-edge models function as “black boxes” offering little transparency in their decision-making mechanisms. The absence of interpretability limits their use in therapeutic environments where confidence, elucidation, and regulatory adherence are essential. There is an increasing need for intelligent systems that attain high diagnostic precision while offering interpretable insights to facilitate clinical decision-making. This paper presents an innovative explainable artificial intelligence framework that incorporates a stacked ensemble deep learning architecture for the classification and segmentation of brain tumors, aiming to bridge the performance–interpretability gap. The proposed approach integrates three complimentary convolutional neural networks—DenseNet121, VGG16, and MobileNetV2—each meticulously fine-tuned on a curated MRI dataset including glioma, meningioma, pituitary tumors, and healthy controls to exploit varied feature extraction capabilities. A logistic regression meta-model consolidates the predictions of various basis classifiers, enhancing generalization, robustness, and transparency. The Local Interpretable Model-agnostic Explanations approach produces case-specific visual explanations that emphasize the picture areas affecting classification results, hence improving interpretability. Analysis of over 7000 MRI images indicates that the suggested framework attains around 99% accuracy, exhibiting good precision, recall, and F1-score across tumor classifications. ROC and precision–recall evaluations validate robust discriminative performance and dependability, underscoring its potential for elucidative clinical decision assistance.</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

EASE-Net: An Explainable AI-Based Segmentation and Ensemble Network for Interpretable Brain Tumor Diagnosis and Classification in MRI Images

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Pirishita Tuteja,
  • Muzafar Ahmad Wani,
  • Niyaz Ahmad Wani,
  • Jatin Bedi

摘要

The identification and categorization of brain tumors from magnetic resonance imaging (MRI) scans pose a significant challenge in neuro-oncology, where prompt and precise diagnosis is crucial for enhancing patient survival and informing treatment approaches. Recent advancements in deep learning have markedly improved the precision and efficacy of automated tumor diagnosis; nonetheless, several cutting-edge models function as “black boxes” offering little transparency in their decision-making mechanisms. The absence of interpretability limits their use in therapeutic environments where confidence, elucidation, and regulatory adherence are essential. There is an increasing need for intelligent systems that attain high diagnostic precision while offering interpretable insights to facilitate clinical decision-making. This paper presents an innovative explainable artificial intelligence framework that incorporates a stacked ensemble deep learning architecture for the classification and segmentation of brain tumors, aiming to bridge the performance–interpretability gap. The proposed approach integrates three complimentary convolutional neural networks—DenseNet121, VGG16, and MobileNetV2—each meticulously fine-tuned on a curated MRI dataset including glioma, meningioma, pituitary tumors, and healthy controls to exploit varied feature extraction capabilities. A logistic regression meta-model consolidates the predictions of various basis classifiers, enhancing generalization, robustness, and transparency. The Local Interpretable Model-agnostic Explanations approach produces case-specific visual explanations that emphasize the picture areas affecting classification results, hence improving interpretability. Analysis of over 7000 MRI images indicates that the suggested framework attains around 99% accuracy, exhibiting good precision, recall, and F1-score across tumor classifications. ROC and precision–recall evaluations validate robust discriminative performance and dependability, underscoring its potential for elucidative clinical decision assistance.