What question did this study set out to answer?

The aim is to create a deep learning-based system to assess patient positioning in chest X-rays, ultimately improving diagnostics.

March 21, 2026BMC Medical Imaging0 citationsOpen Access

Application of deep learning strategies in the standardization and diagnostic efficiency enhancement of chest X-ray imaging

YWYung-Cheng Wang WCWei-Chi Chen KLKang-Ping Lin

Key Points

The aim is to create a deep learning-based system to assess patient positioning in chest X-rays, ultimately improving diagnostics.
Collected and annotated 3663 posterior-anterior chest X-rays for key structural compliance.
Utilized a hybrid deep learning model with Faster R-CNN for localization and ResNet-50 for classification.
Employed data augmentation and progressive training to address class imbalances.
Measured model performance through accuracy, sensitivity, specificity, and Intersection over Union metrics.
Cervical spine compliance achieved 94.54% accuracy, with high sensitivity and specificity.
Costophrenic angle assessment reached 98.06% accuracy, showing remarkable sensitivity.
Scapulae localization was successful in 99.38% of cases, demonstrating strong performance.
Overall, the system offers reproducible evaluations, reducing observer variability in diagnostics.

Abstract

This study aimed to develop a deep learning-based system to automatically assess patient positioning compliance in chest radiographs, focusing on cervical spine visibility, scapular overlap, and costophrenic angle completeness, to improve image quality and diagnostic accuracy. A total of 3663 de-identified posterior-anterior chest X-rays were collected and annotated for cervical spine, scapulae, and costophrenic angles. Compliance criteria included visibility of at least three cervical vertebrae, non-overlapping scapulae, and fully visible costophrenic angles. A hybrid deep learning framework combining Faster R-CNN for region localization and ResNet-50 for compliance classification was employed. Progressive training and data augmentation addressed class imbalance, and model performance was evaluated using accuracy, sensitivity, specificity, Intersection over Union (IoU), and Y-axis IoU (Y-IoU). Cervical spine compliance was identified with 94.54% accuracy, 96.47% sensitivity, and 91.24% specificity. Costophrenic angle assessment achieved 98.06% accuracy, 98.28% sensitivity, and 95.24% specificity, with validation on additional datasets yielding 98.91% accuracy. Scapulae localization succeeded in 99.38% of images, and compliance classification reached 92.10% (left) and 92.38% (right) accuracy, with errors mainly due to patient positioning or slight scapula-lung overlap. The proposed deep learning system provides objective, reproducible evaluation of chest X-ray positioning, enhancing image quality and reducing observer variability. Integration into clinical workflows could support radiographers in achieving standardized radiographs, improving diagnostic reliability and patient care. The AI system developed in this study helps standardize and improve the quality of chest X-ray imaging procedures, reduces errors from manual interpretation, enhances diagnostic accuracy, and serves as an important tool for clinical quality control and professional development of radiologic technologists. It also has the potential for future implementation in multi-center clinical practice.

KI fragen

Bookmark

View Full Paper