Scholarlink

<p>In this paper, we propose strategies for training stronger feature representations and more effective classifiers for chest X-ray (CXR) image analysis by combining attention-enhanced convolutional neural networks (CNNs), self-supervised feature learning, and nonlinear classification. Modern deep learning models for medical imaging often require large labeled datasets and often fail to effectively capture long-range spatial relationships, which are essential for identifying diffuse or widespread abnormalities. To overcome these limitations, we first introduce an attention-augmented DenseNet (DNet-nSA), which embeds self-attention blocks into the DenseNet architecture, allowing the network to capture non-local interactions and highlight clinically important regions. Moreover, we propose replacing the softmax head in DenseNet with a nonlinear classifier, such as a Support Vector Machines (SVM), to further improve the accuracy of CXR image classification. In parallel, we propose a self-supervised multi-classifier framework (SSL-mC), in which feature representations or prediction scores obtained from fine-tuned self-supervised models are fused and subsequently classified using nonlinear learners. This design enhances representation quality and exploits complementary information across models. Experiments on CXR datasets show that both strategies yield substantial improvements: SSL-mC achieves an accuracy of 88%, while DNet-nSA combined with nonlinear classifiers reaches 90.4%. Compared to the baselines, SSL-mC improves accuracy by up to 2.2%, whereas DNet-nSA with SVM achieves improvements of up to 7.1%. These results demonstrate that improving representation learning, either through self-supervision or attention-enhanced CNNs, combined with nonlinear classification leads to more accurate and robust CXR image classification.</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

Training Better Representations and Classifiers for X-Ray Images via Self-Supervised Learning and Attention-Enhanced DenseNet

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Tri-Thuc Vo,
  • Thanh-Nghi Do

摘要

In this paper, we propose strategies for training stronger feature representations and more effective classifiers for chest X-ray (CXR) image analysis by combining attention-enhanced convolutional neural networks (CNNs), self-supervised feature learning, and nonlinear classification. Modern deep learning models for medical imaging often require large labeled datasets and often fail to effectively capture long-range spatial relationships, which are essential for identifying diffuse or widespread abnormalities. To overcome these limitations, we first introduce an attention-augmented DenseNet (DNet-nSA), which embeds self-attention blocks into the DenseNet architecture, allowing the network to capture non-local interactions and highlight clinically important regions. Moreover, we propose replacing the softmax head in DenseNet with a nonlinear classifier, such as a Support Vector Machines (SVM), to further improve the accuracy of CXR image classification. In parallel, we propose a self-supervised multi-classifier framework (SSL-mC), in which feature representations or prediction scores obtained from fine-tuned self-supervised models are fused and subsequently classified using nonlinear learners. This design enhances representation quality and exploits complementary information across models. Experiments on CXR datasets show that both strategies yield substantial improvements: SSL-mC achieves an accuracy of 88%, while DNet-nSA combined with nonlinear classifiers reaches 90.4%. Compared to the baselines, SSL-mC improves accuracy by up to 2.2%, whereas DNet-nSA with SVM achieves improvements of up to 7.1%. These results demonstrate that improving representation learning, either through self-supervision or attention-enhanced CNNs, combined with nonlinear classification leads to more accurate and robust CXR image classification.