Scholarlink

We present a shape-aware thoracic edge map analysis method for pulmonary abnormality screening in chest x-rays (CXR). In CXR, lung abnormalities are often identified by analyzing texture, such as abnormal lung markings, including extra lines (seen in interstitial patterns) and cloud-like opacities (such as consolidations). Tuberculosis (TB) follows a similar pattern, and since an edge map is the derivative of texture, this paper captures complex thoracic edge maps and their spatial relationships (using a log-polar histogram) through shape context (SC). Using SC features, we classify abnormal CXRs with a support vector machine (SVM) model, employing an RBF kernel. On two publicly available datasets, Shenzhen (and Montgomery County), our approach achieves an accuracy of 0.864 (and 0.857), precision of (0.87) (and 0.91), recall of 0.86 (and 0.80), F1 score of 0.86 (and 0.80), and an AUC of 0.925 (0.855). Although the results are comparable, our study suggests that the independent use of edge maps alone may not suffice; they need to be combined with additional advanced shape or texture features for optimal performance.

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

Shape-Aware Thoracic Edge Map Chest X-Ray Representation for Pulmonary Abnormality Screening

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sandeep Chataut,
  • Aashish Ghimire,
  • Anjali Thakur,
  • Longwei Wang,
  • KC Santosh

摘要

We present a shape-aware thoracic edge map analysis method for pulmonary abnormality screening in chest x-rays (CXR). In CXR, lung abnormalities are often identified by analyzing texture, such as abnormal lung markings, including extra lines (seen in interstitial patterns) and cloud-like opacities (such as consolidations). Tuberculosis (TB) follows a similar pattern, and since an edge map is the derivative of texture, this paper captures complex thoracic edge maps and their spatial relationships (using a log-polar histogram) through shape context (SC). Using SC features, we classify abnormal CXRs with a support vector machine (SVM) model, employing an RBF kernel. On two publicly available datasets, Shenzhen (and Montgomery County), our approach achieves an accuracy of 0.864 (and 0.857), precision of (0.87) (and 0.91), recall of 0.86 (and 0.80), F1 score of 0.86 (and 0.80), and an AUC of 0.925 (0.855). Although the results are comparable, our study suggests that the independent use of edge maps alone may not suffice; they need to be combined with additional advanced shape or texture features for optimal performance.