Scholarlink

<p>This paper introduces a hybrid deep learning model combining ConvNeXt and Swin Transformer for classifying brain tumors from MRI scans. The ConvNeXt backbone is employed to obtain detailed local spatial features, whereas the Swin Transformer identifies hierarchical long-range dependencies, facilitating complementary feature representation. The proposed model is evaluated on a combined public MRI dataset of 7,023 images distributed across four categories: glioma, meningioma, pituitary, and no tumor. Experimental results demonstrate that the proposed hybrid architecture outperforms several state-of-the-art convolutional and transformer-based models, achieving an accuracy of 95.37% with competitive precision and F-score. Additionally, qualitative explainability assessment using attention-based visualization techniques offers insight into the model’s decision-making by highlighting diagnostically significant regions. Furthermore, we evaluate the proposed H-ConvNeXt-Swin model on the unified dataset and also report source-stratified performance on each of the three constituent datasets: Figshare, SARTAJ, and Br35H. Future work will focus on validating the proposed framework on multi-center clinical datasets and extending it to more complex tasks such as tumor localization and segmentation.</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

A novel hybrid transformer-based framework (H-ConvNeXt–Swin) to classify brain tumors using MRI

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Essam Abdellatef,
  • Rasha M. Al-Makhlasawy,
  • Nesma Abd El-Mawla,
  • Wafaa A. Shalaby

摘要

This paper introduces a hybrid deep learning model combining ConvNeXt and Swin Transformer for classifying brain tumors from MRI scans. The ConvNeXt backbone is employed to obtain detailed local spatial features, whereas the Swin Transformer identifies hierarchical long-range dependencies, facilitating complementary feature representation. The proposed model is evaluated on a combined public MRI dataset of 7,023 images distributed across four categories: glioma, meningioma, pituitary, and no tumor. Experimental results demonstrate that the proposed hybrid architecture outperforms several state-of-the-art convolutional and transformer-based models, achieving an accuracy of 95.37% with competitive precision and F-score. Additionally, qualitative explainability assessment using attention-based visualization techniques offers insight into the model’s decision-making by highlighting diagnostically significant regions. Furthermore, we evaluate the proposed H-ConvNeXt-Swin model on the unified dataset and also report source-stratified performance on each of the three constituent datasets: Figshare, SARTAJ, and Br35H. Future work will focus on validating the proposed framework on multi-center clinical datasets and extending it to more complex tasks such as tumor localization and segmentation.