MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Quantification of Pixel-Wise Uncertainty Associated with Automatic Segmentation

D Ruan*, T Zhao , D Low , M Steinberg , UCLA, Los Angeles, CA

Presentations

(Tuesday, 7/16/2019) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 6

Purpose: Segmentation approaches have advanced greatly. However, clinical value and translation calls for confidence quantification, which is largely missing. Furthermore, knowledge of uncertainty helps to optimize editing efforts. We have developed a systematic approach to generate simultaneously a contour estimator and its associated uncertainty, on a per-pixel basis.

Methods: We pose the contour and uncertainty estimation problem by putting each image in the context of other (image, contour) elements in a learning cohort. The cohort could be from a public database of similar task or could be from a contour registry to be adjudicated. For each “target� (image, contour) pair to be adjudicated, we have developed a metric-learning method to identify a relevant subset of the cohort that contributes to the task of estimating the true contour for the “target� image and assessing the uncertainty of the given “target� contour. In doing so, we succeed in generating a posterior probability distribution of structure labels from a single contour sample on a given image set. We further quantify the pixel-wise uncertainty using the approximated predictive entropy, to provide a spatially varying quality assessment of the segmentation result. This result generalizes the one-quality-value-per-contour analysis and facilitates optimization of localized editing effort.

Results: We tested the performance of the proposed approach on an MRI-based corpus callosum segmentation task (DSC = 0.95) and a liver CT segmentation (DSC = 0.86). In both, the segmentation results achieve state-of-the-art performance. Furthermore, the uncertainty estimate agrees with human observer impressions, demonstrating its potential value to be used as a QA guidance and editing support module.

Conclusion: Framing the automatic segmentation problem with an underlying statistical backbone and solving it with advanced learning approaches, allows us to benefit from both interpretation and performance gain. This work is currently being extended to CT-based contouring tasks.

Keywords

Not Applicable / None Entered.

Taxonomy

TH- Dataset analysis/biomathematics: Informatics

Contact Email