[Evaluation of Radiograph Accuracy in Skull X-ray Images Using Deep Learning].

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiography

Background:

• Accurate patient positioning is crucial for diagnostic radiography and essential for reproducible follow-up imaging.
• Current methods for evaluating radiograph quality rely on subjective visual assessment by technologists, leading to inter-observer variability.
• Standardized criteria for image acceptance are lacking, impacting the consistency of radiological examinations.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (DCNN) for automated image quality assessment of skull X-ray images.
• To determine the DCNN's ability to objectively classify projection angles and identify radiographs requiring retakes.
• To reduce subjectivity and improve consistency in radiographic image evaluation.

Main Methods:

• Skull X-ray images from 5 phantoms were analyzed using simple and VGG16 DCNN architectures.
• The DCNN models were trained to discriminate between correct and incorrect X-ray projection angles and to identify images needing retakes.
• Model performance was assessed using 5-fold cross-validation and leave-one-out cross-validation with varying input image sizes.

Main Results:

• A simple DCNN architecture with small input image sizes achieved a classification accuracy of 99.75% via 5-fold cross-validation.
• VGG16 achieved 80.00% accuracy with small input image sizes.
• Both architectures showed similar accuracy (around 80%) when using general image sizes.

Conclusions:

• A shallow DCNN combined with small input image sizes is highly effective for classifying four categories of X-ray projection angles, reaching 99.75% accuracy.
• The proposed DCNN method demonstrates potential for automatically detecting subtle projection errors and identifying images that do not meet acceptance criteria.
• This automated evaluation system can provide objective feedback for retakes, potentially reducing radiation exposure and mitigating subjective variations in image quality assessment.