Scholarlink

<p>Accurate and early identification of brain tumors is vital for clinical decision making, yet the heterogeneous appearance of tumors on MRI poses a persistent challenge for automated classification. We propose a deep learning framework that integrates transfer learning, wavelet domain adaptation, attention mechanisms, and a composite multi-loss function. Built on a ResNet50 backbone, the model leverages wavelet decomposition to enhance edge and texture preservation, CBAM to emphasize salient tumor regions, and a hybrid loss (cross-entropy, focal, dice) to address class imbalance. A two-stage training strategy is employed: Stage 1 pretrains the network on the BraTS18 dataset to capture structural patterns, while Stage 2 fine-tunes and serves as the primary evaluation on a dedicated 7023 MRI dataset spanning glioma, meningioma, pituitary tumor, and healthy cases. Our approach achieves 98.44% classification accuracy, outperforming state-of-the-art methods in both accuracy and robustness.</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

Multi-stage wavelet-attention deep learning for brain MRI classification

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Tuan Khoi Tran,
  • Soo-Hyung Kim,
  • Hyung-Jeong Yang,
  • Myungeun Lee

摘要

Accurate and early identification of brain tumors is vital for clinical decision making, yet the heterogeneous appearance of tumors on MRI poses a persistent challenge for automated classification. We propose a deep learning framework that integrates transfer learning, wavelet domain adaptation, attention mechanisms, and a composite multi-loss function. Built on a ResNet50 backbone, the model leverages wavelet decomposition to enhance edge and texture preservation, CBAM to emphasize salient tumor regions, and a hybrid loss (cross-entropy, focal, dice) to address class imbalance. A two-stage training strategy is employed: Stage 1 pretrains the network on the BraTS18 dataset to capture structural patterns, while Stage 2 fine-tunes and serves as the primary evaluation on a dedicated 7023 MRI dataset spanning glioma, meningioma, pituitary tumor, and healthy cases. Our approach achieves 98.44% classification accuracy, outperforming state-of-the-art methods in both accuracy and robustness.