A semi-supervised learning-based quality evaluation system for digital chest radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiography Quality Assessment

Background:

• Digital radiography is essential for disease diagnosis, but manual quality evaluation is unreliable.
• Manual assessment is time-consuming, labor-intensive, and prone to interobserver variability.
• Developing automated, quantitative methods is crucial for efficient radiograph quality evaluation.

Purpose of the Study:

• To develop a rapid and reliable quality evaluation system for digital radiographs.
• To reduce the workload of radiographic technologists.
• To establish quantitative standards for radiograph quality assessment using deep learning.

Main Methods:

• A deep learning system was developed for evaluating frontal chest radiograph quality.
• π-ResUNet was used for semantic segmentation of lung, scapula, and clavicle for patient positioning assessment.
• FasterRCNN was employed for foreign body detection.
• A semi-supervised learning (SSL) strategy with consistency loss was implemented to enhance network performance using unlabeled radiographs.
• Performance was compared against a fully supervised learning (FSL) strategy.

Main Results:

• The SSL-trained network achieved high Dice similarity coefficients (DSC) for segmentation: 0.96 (lung), 0.88 (scapula), and 0.88 (clavicle), outperforming FSL.
• For foreign body detection, the SSL method yielded superior results with an Area Under the ROC Curve (AUC) of 0.90 and Free-response ROC (FROC) of 0.77.
• The proposed system evaluates radiograph quality in under 1 second.

Conclusions:

• The proposed deep learning system effectively evaluates digital radiograph quality.
• Semi-supervised learning significantly improves the performance of the quality evaluation networks.
• The system offers a fast and precise tool for assessing patient positioning and detecting foreign bodies in chest radiographs.