Click here to


Are you sure ?

Yes, do it No, cancel

Spatial Correlation of Radiomics Features with Segmentation Errors of PET-Based Tumor Contours in the Lung

F Yang1*, (1) University of Miami, Miami, FL


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Uncertainty and variability incurring in PET-based lesion target definition have been well recognized; however, the underlying causes remain largely unaddressed. The aim of the current study was to assess if local fine-grained imaging features occupies a role in radiologic image perception and interpretation of PET imaged lung lesions.

Methods: Imaging data employed in the study comprised 26 synthetic PET scans created through the use of an anthropomorphic phantom in conjunction with Monte Carlo simulation. Each dataset featured one PET-positive lesion which varied in terms of shape, heterogeneity, and location inside the lung. Target contours were provided by 10 physicians and a zone of delineation error was created for each lesion through consolidating the volumetric difference of each manual contour in regard to its ground truth. Within the zone of delineation error for each of the simulated lesions, parametric images of 50 radiomics features were generated along with an error rate map, which characterized how frequently manual contouring misclassifies the target on a voxel basis, being created. Spatial correlation between parametric images and error rate maps were assessed using spatial efficiency metric (SPAEF) with the ideal value being 1.0.

Results: Though spatial correlation between the derived parametric images and error rate maps is greatly feature-dependent and could vary substantially, a number of imaging features including gray-level co-occurrence matrix-based contrast and dissimilarity along with gray-level size zone matrix-based short zone emphasis and high intensity short zone emphasis amongst other features showed consistent spatial agreement in regard to the error rate maps with average SPAEF scores above 0.75 among the lesions being examined.

Conclusion: It was demonstrated that likelihood of manual misclassification at the voxel level correlates with certain local fine-grained imaging features. This may further the understanding as to what causes variation in the contouring of PET positive lung lesions.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email