MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Statistical Shape Models for the Automated Detection of OAR Segmentation Abnormalities

E Schreibmann*, H Shim , T Liu , N Esiashvili, Emory University, Atlanta, GA

Presentations

(Wednesday, 8/1/2018) 4:30 PM - 6:00 PM

Room: Karl Dean Ballroom A1

Purpose: Verification of organ as risk (OAR) segmentation quality is time-consuming as a structure’s shape can’t be easily assessed in planar views by an observer. We have developed an artificial intelligence algorithm that compares segmentations against a statistical model of shape variability. Here we report its clinical application.

Methods: The automation is implemented through a software module integrated with the treatment planning through scripting to create reports of segmentation quality. The core of this approach is learning of normal shape appearances quantified as statistical variations from a reference shape. These variations are quantified through principal component analysis to obtain a low-dimensional representation of allowable shape variations. Any deviations from the norm are automatically flagged for review. The first components of the fitted PCA vector are used to mark abnormal shapes and suggested corrections as inferred from the shape model and displayed in modern software to visualize suggested changes with the corresponding images.

Results: The algorithm was applied to plans of 2029 patient across a wide variety of treatment techniques. A range of issues were flagged. Out of 747 brainstem segmentations, 269 (36%) were flagged with minor changes, 2 had segmentations in a few planes and 2 had included other structures in their segmentation. The mandible was the organ with the highest number of inadvertent inclusion of non-organ voxels 58 (21.8%) due to the usage of automated thresholding algorithms. Frequent minor corrections were suggested for eyes (70 cases, 15.3%) and lenses (182, 15.7%). Largest deviations from the expected shape as measured with the Hausdorff distance occurred at 14.2 mm for kidney and 12.1 mm for liver.

Conclusion: We have implemented the automation software that provides objective analysis of OAR segmentation across a multitude of organ shapes and sizes. This verification is valuable as we continue to improve the quality of radiotherapy.

Keywords

Not Applicable / None Entered.

Taxonomy

TH- Dataset analysis/biomathematics: Machine learning techniques

Contact Email