Detection of phase-binning and interpolation artifacts in 4-dimensional computed tomography imaging using deep...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Four-dimensional computed tomography (4DCT) is vital for lung cancer radiotherapy planning, enabling functional avoidance strategies.
• Acquisition artifacts in 4DCT scans compromise downstream analyses like CT-ventilation and dose accumulation.
• Artifacts hinder accurate lung segmentation and deformable image registration, impacting treatment efficacy.

Purpose of the Study:

• To develop 3D deep learning models for voxel-level phase-binning artifact detection in 4DCT images.
• To create a rule-based method for identifying interpolation artifacts within 4DCT scans.
• To enhance the reliability of 4DCT imaging for lung cancer radiotherapy.

Main Methods:

• Generated synthetic 4DCT data with phase-binning and interpolation artifacts for training deep learning models (nnUNet, SwinUNETR).
• Utilized nine diverse clinical 4DCT datasets to ensure model robustness across various artifact severities and patient geometries.
• Developed a localized artifact correction method by replacing artifact-affected voxels with average surrounding lung intensity.

Main Results:

• nnUNet and SwinUNETR models achieved state-of-the-art artifact detection accuracy (average 0.957), with nnUNet reaching 0.965 accuracy.
• The rule-based interpolation detection method demonstrated high performance (0.97 accuracy, sensitivity, and specificity).
• Artifact correction improved lung segmentation mask quality, with 65% achieving Dice scores > 0.95 post-correction.

Conclusions:

• 3D deep learning and rule-based methods provide accurate artifact detection in 4DCT, trained on synthetic and real data.
• The developed models offer interpretability, highlighting artifact-affected voxels for targeted correction.
• SwinUNETR's accuracy and speed show potential for real-time artifact correction guidance or re-scan signaling.