Deep-learning detection of phase-binning artifacts in 4DCT images

4DCT imaging is crucial for precise radiation therapy in lung cancer patients, enabling functional avoidance planning to mitigate radiotherapy toxicity. However, 4DCT scans are frequently impaired by motion-based artifacts that produce abnormalities and corrupt downstream analyses, like lung segmentation masks. This study develops 3D deep learning models to identify the voxels of artifacts within 4DCT images. We introduce a generator to systematically insert synthetic artifacts into artifact-free end-inspiration phases from 9 clinical datasets to ensure robustness across artifact severity and various lung geometries and cancer progressions. As a supervised task, we train modified nnUNet and SwinUNETR frameworks on synthetic 4DCT images. After comprehensive training and testing, both frameworks outperform the state-of-the-art detectors, where nnUNet trained on custom region-based masks achieves an averaged accuracy of 0.995, sensitivity of 0.955, and specificity of 0.998 when inferring artifact-affected axial slices. We propose a localized artifact correction method that simply replaces the predicted artifact-affected voxels with the average lung intensity value, resulting in 74% of cases with Dice scores greater than 0.95 on an evaluation mask (opposed to 9.6% of cases before correction). Despite lower performance on synthetic artifacts, SwinUNETR configurations tend to be more generalizable to true artifacts and demonstrate impressive performance on 310 cases compared to nnUNet configurations. This study reduces the burden of human assessments that are time-consuming and subjective when labeling slices as artifact-free or not, while contributing to the interpretability of AI in medical applications by highlighting which voxels contribute to artifact-affected slices. The accurate and quick SwinUNETR models have the potential to pave the way for more targeted artifact correction methods or signal an imaging technologist in real-time for when to re-scan a patient.