Dead detector element detection in flat panels using convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Quality Assessment

Background:

• Independent testing of medical imaging systems is crucial for unbiased performance evaluation.
• Vendor-reported data on dead detector elements is often incomplete and difficult to verify independently.
• Dead detector elements can significantly impact diagnostic image quality.

Purpose of the Study:

• To develop a vendor-independent method for detecting dead detector elements using convolutional neural networks (CNNs).
• To enable accurate assessment of detector health and its impact on image quality.
• To provide a tool for quality assurance in medical diagnostic imaging.

Main Methods:

• A dataset of 61 flat field images from Varian on-board imaging (OBI) systems was used.
• Convolutional neural networks (CNNs) were trained and validated to identify dead detector elements from subimages.
• Dead pixel maps were acquired to serve as ground truth for training the CNNs.

Main Results:

• CNNs achieved an average precision of 0.96, recall of 0.48, and F1 score of 0.55.
• Performance was limited by data scarcity and imbalance, common in training neural networks with flat-field images.
• A significant performance improvement was observed with a high dynamic range data subset.

Conclusions:

• The study demonstrates the feasibility of using CNNs to detect dead detector elements from flat field images.
• A generalized model could offer independent evaluation of detector health across various vendors and models.
• This method can enhance quality assurance procedures for X-ray, mammography, and fluoroscopy systems.