Binary classification of dead detector elements in flat panel detectors using convolutional neural networks.

Area of Science:

• Medical Physics
• Radiological Imaging Technology
• Computer Vision

Background:

• Digital radiography systems utilize flat panel detectors that can degrade over time due to malfunctioning detector elements.
• These malfunctioning elements, or 'dead pixels,' reduce overall image quality, and their correction data is often proprietary to vendors, limiting physicist access.
• Accurate identification of dead detector elements is crucial for medical physicists performing quality assurance (QA) on these systems.

Purpose of the Study:

• To develop and validate a novel technique for classifying dead detector elements in flat panel detectors at single pixel resolution.
• To assess the generalizability of the developed technique across different detectors, potentially from different vendors.
• To provide medical physicists with a tool for independent quality assurance of digital radiography systems.

Main Methods:

• A novel technique using convolutional neural networks (CNNs) was developed to classify dead detector elements.
• The technique requires acquiring three flat field (noise) images for processing.
• The model's ability to generalize was tested by training on one detector and validating on another, with preprocessing using standard deviation across images.

Main Results:

• Models preprocessed using standard deviation across three images achieved F1 scores from 0.4527 to 0.8107 and recall from 0.5420 to 0.9303.
• Performance was generally better when trained on low exposure datasets compared to high exposure datasets.
• Models using only raw pixel data failed to generalize between detectors, unlike the preprocessed models.

Conclusions:

• CNNs can effectively predict dead detector element maps with single pixel resolution, offering a valuable tool for medical physicists.
• The developed technique demonstrates moderate success in generalizing to different detectors, though further investigation across vendors is needed.
• Acquiring only three flat field images is sufficient for implementing this quality assurance tool, potentially without requiring high exposure settings.