Scholarlink

<p>Accurate detection of brain tumors from MRI images remains a critical challenge due to data imbalance and tumor variability. This research addresses these gaps by proposing and evaluating advanced hybrid and ensemble deep learning models for binary brain tumor classification. A publicly available brain MRI dataset containing 253 images was expanded to 3,720 images using traditional data augmentation techniques and further increased to 5,168 images through Wasserstein Autoencoder (WAE)–based synthetic data generation. This study introduces novel hybrid architectures, such as RSN50-ViTB, and advanced ensemble models, including CVMR-ViT, which integrate CNN, VGG16, VGG19, MobileNetV2, ResNet50V2, ResNet101V2, and Vision Transformer (ViT) architectures to enhance feature learning depth and generalization capability. All models were evaluated across four datasets: the original dataset, the simple augmented dataset, the combined dataset (original + simple augmentation), and the combined dataset (original + WAE augmentation). Across all evaluation settings, the CVMR-ViT Ensemble consistently achieved the best performance. Specifically, on the original dataset, CVMR-ViT attained 95.0% accuracy, 94.9% precision, 94.8% recall, and an F1-score of 94.85%. Performance further improved with simple augmentation, where CVMR-ViT achieved 95.92% accuracy, 96.10% precision, 95.80% recall, and an F1-score of 95.95%. Using the combined dataset (original + simple augmentation), the model reached 97.30% accuracy, 96.50% precision, 95.65% recall, and an F1-score of 97.75%. Finally, with the combined dataset incorporating WAE-generated samples, CVMR-ViT maintained strong and stable performance, achieving approximately 97.0% accuracy, 97.20% precision, 96.85% recall, and an F1-score of 97.02%. These findings demonstrate the superior diagnostic capability of ensemble and hybrid architectures, as well as the critical role of high-quality data augmentation, particularly WAE-based synthesis, in addressing data scarcity and improving model generalization in small medical imaging datasets. Overall, this work advances the state of the art in brain tumor detection and supports the development of reliable AI-assisted diagnostic systems for potential clinical applications.</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

Data-efficient neuroimaging classification via Wasserstein Autoencoder–based augmentation and hybrid deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Priya Mathur,
  • Amit Kumar Gupta,
  • Farhan Sheth

摘要

Accurate detection of brain tumors from MRI images remains a critical challenge due to data imbalance and tumor variability. This research addresses these gaps by proposing and evaluating advanced hybrid and ensemble deep learning models for binary brain tumor classification. A publicly available brain MRI dataset containing 253 images was expanded to 3,720 images using traditional data augmentation techniques and further increased to 5,168 images through Wasserstein Autoencoder (WAE)–based synthetic data generation. This study introduces novel hybrid architectures, such as RSN50-ViTB, and advanced ensemble models, including CVMR-ViT, which integrate CNN, VGG16, VGG19, MobileNetV2, ResNet50V2, ResNet101V2, and Vision Transformer (ViT) architectures to enhance feature learning depth and generalization capability. All models were evaluated across four datasets: the original dataset, the simple augmented dataset, the combined dataset (original + simple augmentation), and the combined dataset (original + WAE augmentation). Across all evaluation settings, the CVMR-ViT Ensemble consistently achieved the best performance. Specifically, on the original dataset, CVMR-ViT attained 95.0% accuracy, 94.9% precision, 94.8% recall, and an F1-score of 94.85%. Performance further improved with simple augmentation, where CVMR-ViT achieved 95.92% accuracy, 96.10% precision, 95.80% recall, and an F1-score of 95.95%. Using the combined dataset (original + simple augmentation), the model reached 97.30% accuracy, 96.50% precision, 95.65% recall, and an F1-score of 97.75%. Finally, with the combined dataset incorporating WAE-generated samples, CVMR-ViT maintained strong and stable performance, achieving approximately 97.0% accuracy, 97.20% precision, 96.85% recall, and an F1-score of 97.02%. These findings demonstrate the superior diagnostic capability of ensemble and hybrid architectures, as well as the critical role of high-quality data augmentation, particularly WAE-based synthesis, in addressing data scarcity and improving model generalization in small medical imaging datasets. Overall, this work advances the state of the art in brain tumor detection and supports the development of reliable AI-assisted diagnostic systems for potential clinical applications.