Scholarlink

<p>The process of segmenting and classifying MRI scans for brain tumor diagnosis and treatment has always been done manually by radiologists. Despite this, since the method is manual, it takes time, is often poorly reproducible, and relies on subjective judgment. Thus, the given paper introduces a novel framework which comprises Swin Transformer for segmentation, a two-level ReLU method to boost features, and an AdaBoost classification layer. Semi-supervised training aims to combine glioma tumor segmentation and classification into a single model to improve performance and handle limited training data. It also applies two methods namely Pseudo-Label Filtering (PLF) and Mean Teacher Learning (MTL), to benefit from unlabeled data without being impacted by noisy labels. The proposed framework is validated and compared with existing state-of-art models based on metrics such as Dice coefficient, Intersection over union (IOU), Accuracy (%), F1-score (%) and Area under curve (AUC). The results show that the proposed framework attains remarkable results as: Dice 98.2%, IoU 94.8%, Accuracy 99.6%, F1-Score 99.25%, and AUC 99.3%.</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

MT-Swin boost: a transformer-based framework for glioma tumor detection with semi-supervised learning and feature boosting

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Abhishek Mishra,
  • Nandita Pradhan

摘要

The process of segmenting and classifying MRI scans for brain tumor diagnosis and treatment has always been done manually by radiologists. Despite this, since the method is manual, it takes time, is often poorly reproducible, and relies on subjective judgment. Thus, the given paper introduces a novel framework which comprises Swin Transformer for segmentation, a two-level ReLU method to boost features, and an AdaBoost classification layer. Semi-supervised training aims to combine glioma tumor segmentation and classification into a single model to improve performance and handle limited training data. It also applies two methods namely Pseudo-Label Filtering (PLF) and Mean Teacher Learning (MTL), to benefit from unlabeled data without being impacted by noisy labels. The proposed framework is validated and compared with existing state-of-art models based on metrics such as Dice coefficient, Intersection over union (IOU), Accuracy (%), F1-score (%) and Area under curve (AUC). The results show that the proposed framework attains remarkable results as: Dice 98.2%, IoU 94.8%, Accuracy 99.6%, F1-Score 99.25%, and AUC 99.3%.