Using deep learning for pixel-defect corrections in flat-panel radiography imaging.

Area of Science:

• Medical Imaging
• Image Processing
• Deep Learning

Background:

• Flat-panel radiography detectors use thin-film transistor (TFT) panels for high-quality X-ray imaging.
• Pixel defects in TFT panels degrade image quality, reduce production yield, and increase costs.
• Traditional defect correction algorithms are complex and difficult to optimize.

Purpose of the Study:

• To develop and evaluate deep learning-based algorithms for correcting pixel defects in X-ray images.
• To compare the performance of artificial neural networks (ANN), convolutional neural networks (CNN), concatenate CNN, and generative adversarial networks (GAN) for defect correction.

Main Methods:

• Investigated various deep learning models including ANN, CNN, concatenate CNN, and GAN.
• Trained models using chest X-ray images with maximal defect sizes of 2x2 and 3x3 pixels.
• Utilized mean square error (MSE) as the loss function to evaluate correction performance.

Main Results:

• Concatenate CNN achieved the lowest MSE (68.21), indicating superior defect correction.
• ANN demonstrated comparable performance to concatenate CNN with significantly lower encoding complexity (MSE: 69.40).
• All deep learning methods outperformed the conventional template match correction method.

Conclusions:

• Concatenate CNN provides the best defect-correction performance among the tested deep learning models.
• Single-layer ANN offers an efficient alternative, balancing high correction accuracy with reduced computational complexity.
• ANN is a viable solution for real-time defect correction in radiography systems.