Scholarlink

<p>Magnetic Resonance Imaging (MRI) requires high precision in both lesion localization and tumor classification for reliable clinical diagnosis. However, most existing deep learning approaches treat segmentation and classification as independent tasks or rely predominantly on global image features, limiting their ability to exploit tumor-specific structural information. To address these limitations, this paper proposes a Transformer-integrated multistage framework, termed TMSF-BTSC, for unified brain tumor segmentation and classification. The key novelty of the proposed approach lies in the explicit propagation of segmentation masks to guide multistage feature extraction and attention-based classification, thereby establishing a strong coupling between pixel-level localization and image-level decision-making. The framework combines a UNETR-based segmentation backbone with a tumor-aware feature extraction pipeline and a structurally aligned attention-driven classification head, enabling effective suppression of background interference and enhancement of lesion-centric representations. Experimental evaluation on a BRISC2025-like multi-class MRI dataset demonstrates that the proposed model achieves a classification accuracy of 93.56%, precision of 0.8726, recall of 0.8724, and F1-score of 0.8722. The model attains a Matthews Correlation Coefficient (MCC) of 0.8291, indicating strong performance under class imbalance, along with high class-wise separability reflected by a maximum ROC-AUC of 0.9940 and a macro-average AUC of 0.9773. Comparative analysis against five state-of-the-art models confirms consistent performance improvements, while ablation studies validate the contribution of each architectural component. Despite its multistage design, the framework maintains moderate computational complexity (31.2&#xa0;million parameters) and low inference latency (21.3ms per image), highlighting its practical applicability. Overall, the proposed segmentation-directed, tumor-aware learning strategy offers a robust and efficient solution for automated brain tumor diagnosis.</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

Transformer-Integrated Multistage Tumor-Aware Framework for Brain Tumor Segmentation and Classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月23日

查看全文
  • K. P. Linija,
  • S. Rajesh

摘要

Magnetic Resonance Imaging (MRI) requires high precision in both lesion localization and tumor classification for reliable clinical diagnosis. However, most existing deep learning approaches treat segmentation and classification as independent tasks or rely predominantly on global image features, limiting their ability to exploit tumor-specific structural information. To address these limitations, this paper proposes a Transformer-integrated multistage framework, termed TMSF-BTSC, for unified brain tumor segmentation and classification. The key novelty of the proposed approach lies in the explicit propagation of segmentation masks to guide multistage feature extraction and attention-based classification, thereby establishing a strong coupling between pixel-level localization and image-level decision-making. The framework combines a UNETR-based segmentation backbone with a tumor-aware feature extraction pipeline and a structurally aligned attention-driven classification head, enabling effective suppression of background interference and enhancement of lesion-centric representations. Experimental evaluation on a BRISC2025-like multi-class MRI dataset demonstrates that the proposed model achieves a classification accuracy of 93.56%, precision of 0.8726, recall of 0.8724, and F1-score of 0.8722. The model attains a Matthews Correlation Coefficient (MCC) of 0.8291, indicating strong performance under class imbalance, along with high class-wise separability reflected by a maximum ROC-AUC of 0.9940 and a macro-average AUC of 0.9773. Comparative analysis against five state-of-the-art models confirms consistent performance improvements, while ablation studies validate the contribution of each architectural component. Despite its multistage design, the framework maintains moderate computational complexity (31.2 million parameters) and low inference latency (21.3ms per image), highlighting its practical applicability. Overall, the proposed segmentation-directed, tumor-aware learning strategy offers a robust and efficient solution for automated brain tumor diagnosis.