Scholarlink

To improve treatment plans and patient outcomes, brain tumors must be accurately and promptly identified. This research introduces a unified and interpretable deep learning model aimed at classifying brain tumors using MRI images. The study leverages a structured dataset comprising 7023 MRI scans, distributed across four distinct categories: glioma, meningioma, pituitary tumor, and healthy (no tumor) cases. An entirely custom convolutional neural network achieved an 85% validation accuracy. Additionally, three models for transfer learning–DenseNet121, VGG16, and ResNet50–were put into practice and evaluated in comparison. Without the need for base layer fine-tuning, VGG16 produced an accuracy of 88%, ResNet50 84%, and DenseNet121 92%. The proposed framework incorporates several diagnostic visualization techniques to improve interpretability and clinical relevance: hemispheric dominance analysis using region-wise activation segmentation, severity classification based on spatial activation area thresholds, tumor localization using Grad-CAM heatmaps, and a novel fusion-based severity scoring mechanism that combines grayscale pixel distribution and Grad-CAM activation intensity. In order to investigate model limitations, a thorough misclassification analysis was carried out with the use of visual overlays. To validate the effectiveness of the models, a diverse set of evaluation tools was employed, including class-wise confusion matrices, ROC-AUC analysis, and statistical indicators such as precision, recall, and the F1 metric. This study offers a useful and reliable tool to help radiologists with real-time brain tumor detection and clinical decision-making by improving interpretability in addition to delivering competitive classification performance.

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 Deep Learning for Brain Tumor Classification and Severity Assessment Using CNN and Transfer Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hosna Ara Begum,
  • Fatema Binte Emran,
  • Shahin Ara Begum

摘要

To improve treatment plans and patient outcomes, brain tumors must be accurately and promptly identified. This research introduces a unified and interpretable deep learning model aimed at classifying brain tumors using MRI images. The study leverages a structured dataset comprising 7023 MRI scans, distributed across four distinct categories: glioma, meningioma, pituitary tumor, and healthy (no tumor) cases. An entirely custom convolutional neural network achieved an 85% validation accuracy. Additionally, three models for transfer learning–DenseNet121, VGG16, and ResNet50–were put into practice and evaluated in comparison. Without the need for base layer fine-tuning, VGG16 produced an accuracy of 88%, ResNet50 84%, and DenseNet121 92%. The proposed framework incorporates several diagnostic visualization techniques to improve interpretability and clinical relevance: hemispheric dominance analysis using region-wise activation segmentation, severity classification based on spatial activation area thresholds, tumor localization using Grad-CAM heatmaps, and a novel fusion-based severity scoring mechanism that combines grayscale pixel distribution and Grad-CAM activation intensity. In order to investigate model limitations, a thorough misclassification analysis was carried out with the use of visual overlays. To validate the effectiveness of the models, a diverse set of evaluation tools was employed, including class-wise confusion matrices, ROC-AUC analysis, and statistical indicators such as precision, recall, and the F1 metric. This study offers a useful and reliable tool to help radiologists with real-time brain tumor detection and clinical decision-making by improving interpretability in addition to delivering competitive classification performance.