Program Information
Automatic Recognition of Streak Artifacts In CT Region of Interests Using Gradient Direction Distribution Method for Radiomics Analysis
L Wei*, B Rosen , A Eisbruch , I El Naqa , University of Michigan, Ann Arbor, MI
Presentations
TU-H-FS4-2 (Tuesday, August 1, 2017) 4:30 PM - 6:00 PM Room: Four Seasons 4
Purpose: Extraction of radiomics features from head and neck CT images is susceptible to dental implant artifacts. This work devises an automatic approach to detect CT metal artifacts in regions of interest (ROIs), e.g. tumor, for efficient and accurate preprocessing of radiomics input images. A new statistical method based on gradient distribution is proposed.
Methods: In a set of 44 head and neck cancer patients, with 10 cases having visually identified streaking artifacts in ROIs, two methods are introduced: the classical total variation approach, and a new gradient-based histogram statistic approach. Considering the characteristic of metal streak artifacts, which contain bright and dark streaks originating from the metal, the total variation should increase in images with artifacts. Thus, using total variation for artifacts is quite straightforward. However, it fails to distinguish artifacts from real textures in ROIs, leading to loss of valuable data for further analysis. Since streak artifacts are directional lines, the distribution difference of the gradient orientations for ROI pixels with and without artifacts makes the pattern recognition accurate and immune to false detection of ROI textures. More specifically, counting how many pixels in ROIs fall into certain number of angle intervals will provide a distribution and those with higher variance correspond to artifacts cases since the existence of directional lines.
Results: Our gradient direction distribution method provided an efficient, simple, and accurate strategy to automatically recognize ROIs affected by streak artifacts. For the 10 patient volumes with streak artifacts (total patient number 44), 8 were successfully detected and gives the accuracy of 95.5%, while total variation gives only 75.0% accuracy.
Conclusion: Although the gradient direction method is applied in the preprocessing step for radiomics studies, the idea possesses the potential to be applied to other image processing and computer vision problems.
Contact Email: