Scholarlink

Purpose <p>This study compares lung nodule detection with nine machine learning models, deep (DL) and non-deep (ML), using the same chest X-ray database.</p> Methods <p>Three Support Vector Machine models, four Random Forest, one SqueezeNet, and one YOLO were trained and validated using the same computing facilities and the same anatomical region as input, consisting of the segmentation of both lungs. The images were further pre-processed by a common algorithm before the ML models and by the same UNet convolutional network before the DL models. The ML models were trained based on a feature vector calculated on the detection of candidate nodules, while the DL models were trained directly from the image segmentation. The best ML model and the DL models were also tested using an external database. The models were compared based on accuracy, precision, sensitivity, and F1 score.&#xa0;</p> Results <p>Under the present standardized experimental conditions, the ML models were found to demonstrate a better balance between metrics, and the best model had generalization power. The DL models showed biases due to overfitting to the training data. The shortcomings of the training conditions for the DL models were analyzed.&#xa0;</p> Conclusion <p>All tested models performed better than humans. ML models performed better under limited training data, and DL showed theoretical potential but requires further optimization.</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

Deep learning vs. machine learning in the detection of pulmonary nodules with chest X-rays

  • 期刊论文

  • 发表日期：

    2025年11月7日

查看全文
  • Marlen Perez-Diaz,
  • Pedro Pablo Camellón,
  • Hanlert Cruz Corzo,
  • Melisa Llody Fajardo,
  • Jhon A. Fernández Castro,
  • Elizabeth Martínez Machado,
  • Yusely Ruiz González

摘要

Purpose

This study compares lung nodule detection with nine machine learning models, deep (DL) and non-deep (ML), using the same chest X-ray database.

Methods

Three Support Vector Machine models, four Random Forest, one SqueezeNet, and one YOLO were trained and validated using the same computing facilities and the same anatomical region as input, consisting of the segmentation of both lungs. The images were further pre-processed by a common algorithm before the ML models and by the same UNet convolutional network before the DL models. The ML models were trained based on a feature vector calculated on the detection of candidate nodules, while the DL models were trained directly from the image segmentation. The best ML model and the DL models were also tested using an external database. The models were compared based on accuracy, precision, sensitivity, and F1 score. 

Results

Under the present standardized experimental conditions, the ML models were found to demonstrate a better balance between metrics, and the best model had generalization power. The DL models showed biases due to overfitting to the training data. The shortcomings of the training conditions for the DL models were analyzed. 

Conclusion

All tested models performed better than humans. ML models performed better under limited training data, and DL showed theoretical potential but requires further optimization.