MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Unsupervised Classification Routine to Correlate Nonlinearly Related Multiple Images: An Example for CT/CBCT Lung Images Normalization

A Chu1*, J Kim2 , Z Xu3 , S Ryu4 , W Liu5 , W Tome6 , (1)(2)(3)(4) Stony Brook University, Stony Brook, NY, (5) Yale Univ. School of Medicine, New Haven, CT, (6) Montefiore Medical Center, Bronx, NY

Presentations

(Sunday, 7/14/2019)  

Room: ePoster Forums

Purpose: Like the density table to CT for radiation dose calculation, this study searches a relation to any CT/CBCT pair for treatment response studies. Unlike the more forgiving HU errors to dosimetric deviation (Yoo, et.al. 2006), the tolerance for response measurement is much less; i.e. the sensitivity of measurement should be much less than the radiological changes. We continued our previous study using piecewise linear CT/CBCT normalization for their nonlinear relationship.

Methods: Last year we presented piecewise linearity can be reached by each clustered intensity ranges and spatial segments. Because the optimization was directly taken on the information of voxel-based intensities with their geometric registration, the process was time consuming. In this study, instead of analyzing all voxels, the intensity classification was handled by some more intensity features with less datapoints; i.e. the curvature of histogram, its 1st, 2nd derivatives, cumulative distribution functions. Instead of using spatial registration, the image pairs were spatially divided into 32 sectors, and the goodness of CT/CBCT pair matching evaluation based on the distribution of distances to the image’s centroid of mass for each sector. Also we developed the rules for consuming all the possibilities finite number for arranging clustered sub-groups classification and recombination to reach the goal of a certain number of linear pieces.

Results: As the improved approach can results in faster optimization, we were able to find out more possibilities to find out better matched CT/CBCT clusters. As suggested in last presentation, a lung image intensity-clustered into 4 segments should be sufficient to be equivalent to 256 gray-levels conventional lung. The process started with over-clustering number on CT (e.g. 10) to be matched into 4~8 clustered CBCT.

Conclusion: The piecewise linear normalization is feasible for continuous, non-linear relationship like CT/CBCT. It could be extended to discontinuous, non-linear relationship like CT/MRI.

Keywords

CT, Cone-beam CT, Image Correlation

Taxonomy

IM- Cone Beam CT: Machine learning, computer vision

Contact Email