Click here to


Are you sure ?

Yes, do it No, cancel

Investigation of CT and CBCT Normalization; An Example of Unsupervised Classification

A Chu1*, J Kim1* , A Hsia1* ,Z Xu1* , (1) Stony Brook University, Stony Brook, NY


(Sunday, 7/29/2018) 1:00 PM - 1:55 PM

Room: Davidson Ballroom B

Purpose: It is convenient to use planning CT with follow-up CBCTs for treatment outcome observations. An appropriate preparation for CT and CBCT is critical for further analyses like machine learning. Here we demonstrated an optimal match can be reached by an unsupervised procedure in spite of significant difference between CT and CBCT.

Methods: An electron density phantom (CIRS) and ten lung patients were studied. (1) Low-noise was removed before a z-scored normalization on each image set. (2) The k-means with the cluster number k (8~16) decomposes an image into k sub-groups. (3) The decomposed k masks were “re-assembled� to match its volume and reach a final number m sub-groups (i.e. m < k). (4) Final match for each pair (histogram) curves of subgroups was based “iterative closest point� (ICP).

Results: (1) The study showed the normalization on CT/CBCT can be used linear procedure as physical density less than about 1.117 g/cm3. A non-linear process would be needed to normalize higher-density contents in CT/CBCT. (2) Choosing lung study is relatively easier for CT/CBCT matching because relatively simple processes is feasible for most lower-density tissue in lung. (3) However, as some decomposed subgroups indicated, artifacts from where the higher-intensity veins distribute degraded the imaging quality. Some degraded areas are irreversible. (4) A lung image intensity sampled by k-means with 8-shaped gray can be comparable to a uniform-scaled 256-shaded of convention because the optimization can recognize the characteristics of population/similarity under the spectrum for a given image. Therefore the success of following classifications was mainly relied on those meaningful sample spots. (5) An example with k=8 and m=2 proved an optimal match can be achieved as their anatomies agree each other, and therefore, the follow sub-group normalizations (ICP) can be reliable.

Conclusion: Lung CT/CBCT can be properly normalized by this procedure.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email