Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session B

Poster F2 - Deep learning detection of motion artifacts in 4DCT images

Thursday, October 24, 2024

2:45 PM - 3:45 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Nathan Feldt

Undergraduate Researcher
The University of Texas at Austin
Seattle, Washington, United States

Primary Investigator(s)

Jorge Cisneros, PhD (he/him/his)

Postdoctoral Fellow
The University of Texas at Austin, United States

Last Author(s)

Edward Castillo, PHD

Associate Professor
University of Texas at Austin, Texas, United States

Introduction: 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 lung abnormalities and corrupt downstream analyses, like lung segmentation and tumor tracking [1]. Previous studies identify artifact-affected stacks of axial slices using normalized cross correlations and differences between adjacent stacks. More recently, with the increase in computational power, further studies have employed deep learning models to detect individual artifact-affected axial slices from 2D coronal slices, achieving the highest accuracy of 0.86 [2]. Despite these recent successes, there is a current lack of precise artifact localization that would not only improve the interpretability of artifact detection methods, but also pave the way for more targeted artifact correction methods. To that end, this study trains state-of-the-art deep learning frameworks, specifically nnUNet and SwinUNETR, to segment both synthetically-generated artifacts and true artifacts found in the end-inspiration 4DCT phase [3][4]. Additionally, we propose a simple localized artifact correction method to improve lung segmentation masks by replacing intensity values of artifact-affected voxels with the surrounding average lung intensity value.

Materials and

Methods: To create a diverse and generalizable training set for our deep learning models, we develop a synthetic artifact generator that mimics the distributions of real artifacts. Figure 2.2 describes how the generator inserts artifacts into an artifact-free end-inspiration phase by randomly selecting stacks from another phase in a patient’s 4DCT set. Due to limited “clean” scans, we repeatedly use them as templates to infuse a variety of artifacts with changes in stack counts, stack height, position, and Gaussian blur. Binary segmentation artifact masks are then computed by thresholding the absolute difference between a clean image and the synthetic artifact-affected image. We trained two deep learning models, nnUNet and SwinUNETR, to detect artifacts in full volumetric 4DCT images using the weighted sum of Binary Cross Entropy (BCE) and Soft Dice Loss for nnUNet and asymmetric Unified Focal Loss for SwinUNETR [5]. To enhance generalization to true artifacts, we randomly augmented the training set by applying rotation, scaling, Gaussian noise, brightness, resolution degradation, gamma correction, and mirroring. With the known location of artifact-affected voxels, the proposed localized artifact correction method essentially camouflages from an automatic segmentation tool the voxels that cause duplicated, overlapping, and cut-off structures. To quantify the method’s performance, we rely on Dice scores and boundary smoothness (ratio of surface area to volume) of the corrected lung masks, which further allows us to verify its success on true artifacts when no ground-truth artifact masks are given via the Two One-Sided Tests for equivalence testing (see Figure 2.4).

Results, Conclusions, and Discussions: After training, the SwinUNETR model outperformed state-of-the-art detectors, including nnUNet, achieving weighted-by-volume averages of 0.989 accuracy, 0.974 sensitivity, and 0.991 specificity at the slice level. After applying the proposed localized correction method on 1255 cases with synthetic artifacts, all but 3 lung segmentations had Dice scores above 0.90 on an evaluation mask, with a mere 1.8% of cases having scores less than 0.95 (opposed to 33.4% of cases pre-correction). Additionally, we show with strong evidence (p > 0.1 in TOST) that the boundary smoothness distribution of 310 corrected true cases with predicted masks is equivalent to the distribution of the corrected synthetic cases with predicted masks, ensuring the DL model is performing well on true artifacts. This further demonstrates SwinUNETR’s ability to generalize to patients’ scans with various cancer progressions and artifact types.
Despite these successes, our DL models struggled to fully segment very large artifacts, most likely because artifacts of these sizes were uncommon in the synthetic training sets. Although the correction method on these artifacts is sub-par, a rescan would most likely already be required. Overall, the accurate and quick SwinUNETR model has 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. Future work will focus on extending these methods to other phases of the breath cycle and developing more advanced artifact correction techniques to further enhance the accuracy and effectiveness of functional avoidance radiotherapy treatment in lung cancer patients.

References
[1]Yamamoto et al., Retrospective analysis of artifacts in four-dimensional CT images of 50 abdominal and thoracic radiotherapy patients, International Journal of Radiation Oncology, Biology, Physics, 2008.
[2]Madesta et al., Deep learning-based conditional inpainting for restoration of artifact-affected 4DCT images, Medical Physics,2023.
[3]Hatamizadeh et al., Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images, International MICCAI Brainlesion Workshop,2021.
[4]Isensee et al., nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation, Nature Methods,2021.
[5]Yeung et al., Unified focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image segmentation, Computerized Medical Imaging and Graphics, 2022.

Acknowledgements (Optional):