Scholarlink

Chest X-rays are the most common tool for diagnosing various thoracic diseases. However, manual interpretation is time-consuming and prone to human error. This paper presents a deep learning approach for automated pathology detection in CXRs using the Customized DenseNet-121 model. The model performs binary classification to identify 14 pathologies, including cardiomegaly, pneumothorax, mass, and edema. To address class imbalance in medical imaging datasets, weight normalization is applied. Additionally, the visualization technique of Grad-CAM enhances interpretability by pointing out the most critical regions influencing the model’s decisions, which helps healthcare practitioners assess. Toward further refinement of segmentation and improvement in precision of localization, we incorporate a customized U-Net model to enhance better delineation of regions of interest. Our model achieves an overall AUC of 87%, showing the highest accuracy. The customized U-Net integration improves segmentation performance, reducing localization error by 15%. This approach not only enhances diagnostic accuracy but also provides transparent decision-making, making it a valuable tool for medical professionals.

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

Anomaly Detection in Lungs Using Deep Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shailaja Uke,
  • Mohit Garg,
  • Suyash Chandolikar,
  • Swayam Chandak,
  • Shriraj Nelekar

摘要

Chest X-rays are the most common tool for diagnosing various thoracic diseases. However, manual interpretation is time-consuming and prone to human error. This paper presents a deep learning approach for automated pathology detection in CXRs using the Customized DenseNet-121 model. The model performs binary classification to identify 14 pathologies, including cardiomegaly, pneumothorax, mass, and edema. To address class imbalance in medical imaging datasets, weight normalization is applied. Additionally, the visualization technique of Grad-CAM enhances interpretability by pointing out the most critical regions influencing the model’s decisions, which helps healthcare practitioners assess. Toward further refinement of segmentation and improvement in precision of localization, we incorporate a customized U-Net model to enhance better delineation of regions of interest. Our model achieves an overall AUC of 87%, showing the highest accuracy. The customized U-Net integration improves segmentation performance, reducing localization error by 15%. This approach not only enhances diagnostic accuracy but also provides transparent decision-making, making it a valuable tool for medical professionals.