Evaluating the Sensitivity of Deep Learning Based Auto-Segmentation to Image Features in Thoracic Segmentation
W Zhu1*, Y Sheng2, Q Chen3, X Li4, D Carpenter5, C Champ6, H Stephens7, Q Wu8, A Diaz9, X Feng10, Q Wu11, (1) Duke University Medical Center, Durham, NC, (2) Duke University Medical Center, Durham, NC, (3) University of Kentucky, Lexington, KY, (4) Duke University Medical Center, Durham, NC, (5) Duke University Medical Center, Durham, NC, (6) Duke University Medical Center, Durham, NC, (7) Duke University, Durham, NC, (8) Duke University Medical Center, Durham, NC, (9) Duke University Medical Center, Durham, NC, (10) University of Virginia, Charlottesville, VA, (11) Duke University Medical Center, Durham, NC
Presentations
(Saturday, 3/26/2022) [Central Time (GMT-5)]
Purpose: To understand current auto-segmentation tools better by identifying limits on model outputs and sensitivity of models to inputs. By quantifying this performance, the impact of features on performance can be identified.
Methods: The INTContour tool developed by Carina was used to analyze images from a dataset of images identified by provider, orientation, laterality, target side, implant, and treatment of regional nodes. Several models were developed to identify features and how they relate to model performance. These models consisted of a laterality and target controlled model for the right breast, a laterality controlled model for the right breast, a target controlled model for the breast planning target volume (PTV), and physician controlled models for right breast PTVs. The physician models were created using data from three physicians, models were created using each individual physician and a combination of all three. After training models, performance of model was assessed by examining segmentation of images outside of the training set and calculating the Dice similarity coefficient (DSC).
Results: Models created without controlling for both laterality and orientation demonstrated contouring of extraneous regions and had poorer performance. Models created without controlling for provider were relatively similar in performance to models that controlled for provider. Models controlled for both laterality and orientation demonstrated more consistent contouring which was reflected by both validation and training DSC values as well as the lack of extraneous contouring.
Conclusion: Our data indicates that controlling for model features like orientation and laterality are useful in eliminating extraneous contouring of regions using the INTContour auto-segmentation tool. Auto-segmentation shows potential in being able to generate relatively similar contours to manually segmented images in a fraction of the time. Further work in post processing can be used to improve performance.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email