Collimation border with U-Net segmentation on chest radiographs compared to radiologists.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Chest radiography (CXR) necessitates radiation dose monitoring and quality assurance (QA).
• Proper collimation is crucial for minimizing patient radiation exposure, adhering to the ALARA principle.
• Automating QA processes can enhance efficiency and accuracy in radiographic procedures.

Purpose of the Study:

• To evaluate the efficacy of U-Net convolutional neural networks (U-CNNs) for automatic lung segmentation in CXRs.
• To determine if U-CNNs can calculate optimized collimation borders for CXR.
• To assess the performance of U-CNNs on a limited CXR dataset for QA applications.

Main Methods:

• Trained and validated three U-CNN models with varying dimensions (128x128, 256x256, 512x512) on 662 CXRs with manual lung segmentations.
• Employed five-fold cross-validation for model assessment.
• Compared U-CNN segmentation and collimation border accuracy against manual segmentations by radiographers and junior radiologists using Dice Scores (DS) and external validation on 50 CXRs.

Main Results:

• U-CNNs achieved high Dice Scores (DS) for lung segmentation (0.93-0.96) and collimation border detection (0.95).
• U-CNN performance closely matched junior radiologists' accuracy (DS 0.97 for both segmentation and collimation).
• One radiographer's manual segmentation differed significantly from U-CNN results (p=0.016).

Conclusions:

• U-CNNs reliably segment lungs and suggest collimation borders with high accuracy on CXRs.
• The developed algorithm demonstrates potential for automating collimation auditing within CXR QA programs.
• Automatic lung segmentation models can generate collimation borders for improved CXR QA.