Scholarlink

<p>Deep learning achieves high accuracy in medical imaging but requires large datasets. Performance decreases in small datasets. This study proposes a deep feature engineering (DFE) framework for calcaneal spur detection in X-ray images. A curated dataset of 775 X-ray images was analyzed, with 316 labeled as no finding and 459 as calcaneal spur. The framework has five phases: (i) feature extraction from 19 pretrained CNNs, (ii) feature selection with iterative neighborhood component analysis (INCA), (iii) classification with a t-algorithm-based k-nearest neighbors (tkNN) ensemble, (iv) generation of voted outcomes through iterative majority voting (IMV), and (v) final selection using a greedy algorithm. The framework achieved 93.42% accuracy. Nineteen CNN outcomes and seventeen IMV-based voted outcomes were evaluated. The greedy step selected the result with the highest accuracy. Spur detection reached higher sensitivity than no finding detection, reflecting the visual similarity between normal heels and early spur cases. The proposed DFE framework attains high accuracy on a small biomedical dataset. Its integration of CNN-based features, INCA, tkNN, IMV, and greedy selection provides a lightweight and generalizable method for medical 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

Calcaneal spur detection from lateral foot radiographs using deep feature engineering

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Sukru Demir,
  • Bugra Can,
  • Omer Faruk Goktas,
  • Mehmet Baygin,
  • Sengul Dogan,
  • Turker Tuncer

摘要

Deep learning achieves high accuracy in medical imaging but requires large datasets. Performance decreases in small datasets. This study proposes a deep feature engineering (DFE) framework for calcaneal spur detection in X-ray images. A curated dataset of 775 X-ray images was analyzed, with 316 labeled as no finding and 459 as calcaneal spur. The framework has five phases: (i) feature extraction from 19 pretrained CNNs, (ii) feature selection with iterative neighborhood component analysis (INCA), (iii) classification with a t-algorithm-based k-nearest neighbors (tkNN) ensemble, (iv) generation of voted outcomes through iterative majority voting (IMV), and (v) final selection using a greedy algorithm. The framework achieved 93.42% accuracy. Nineteen CNN outcomes and seventeen IMV-based voted outcomes were evaluated. The greedy step selected the result with the highest accuracy. Spur detection reached higher sensitivity than no finding detection, reflecting the visual similarity between normal heels and early spur cases. The proposed DFE framework attains high accuracy on a small biomedical dataset. Its integration of CNN-based features, INCA, tkNN, IMV, and greedy selection provides a lightweight and generalizable method for medical image classification.