Click here to

Session: Therapy: Patient Safety and Quality Improvement [Return to Session]

Development of a Quantitative Method to Evaluate Organ Segmentation for Enhanced Error Detection

E Pryser*, F Reynoso, M Schmidt, G Hugo, N Knutson, Washington University School of Medicine, St. Louis, MO

Presentations

TU-IePD-TRACK 6-2 (Tuesday, 7/27/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: Accurate segmentation of organs at risk (OARs) for treatment planning is crucial for the delivery of high-quality care and the prevention of adverse events in radiation therapy. A method to rapidly and quantitatively evaluate the geometric characteristics of an organ against the expected geometry is proposed as a technique for detecting segmentation errors.

Methods: Relationships between organ volume, patient gender, and geometric characteristics including length, width, depth, mean area, and complexity were modeled as a least-squares linear fit for 30 common organs based on metrics gathered from 1,331 patients. Complexity is defined in this context as the ratio of the organ’s volume (Vₒₐᵣ) to that of the bounding-box. For each individual organ, the perpendicular distance from the linear fit with slope M and intercept B was multiplied by a weighting-factor (ω) to generate a sub-score (μ) for each characteristic: μₐᵣₑₐ=ωₐᵣₑₐ*|Mₐᵣₑₐ*Vₒₐᵣ-AREAₒₐᵣ*Bₐᵣₑₐ|*(Mₐᵣₑₐ²+1)⁻². The final score (Sₒₐᵣ) was calculated as the sum of individual sub-scores and a sub-score based on the organ’s volume and the mean organ volume (Vₐᵥ) and standard deviation (σᵥ) determined from an independent dataset: Sₒₐᵣ=μ_area+μ_depth+μ_width+μ_length+ωᵥ*|Vₒₐᵣ-Vₐᵥ|*σᵥ⁻¹. Individual weighting-factors were iteratively adjusted using a subset of the data to optimally assign higher scores to those organs with contouring errors, with a value of Sₒₐᵣ>1 indicating a suspected error.

Results: Of the 10,846 organs in the dataset not used in algorithm training, 746 erroneous contours were correctly identified, 7 erroneous contours were not identified, 2,071 contours were incorrectly identified as erroneous, and the remaining 8,022 were correctly identified as error-free, giving an overall sensitivity of 99.1% and specificity of 79.5%.

Conclusion: A method of quantitatively evaluating OAR contours against demographic data has been developed with the goal of identifying contouring errors. As a classifier, the proposed structure score demonstrates excellent sensitivity, and reasonable specificity in detecting errors in OAR delineation.

Funding Support, Disclosures, and Conflict of Interest: Washington University receives research support from Siemens, Varian Medical Systems, and ViewRay. GDH reports personal fees from Varian Medical Systems outside the scope of the present work. MCS reports personal consulting fees and honoraria with Varian Medical Systems outside of the submitted work.

ePosters

    Keywords

    Segmentation, Shape Analysis, Classifier Design

    Taxonomy

    IM/TH- Image Segmentation Techniques: Clustering, classification and threshold

    Contact Email

    Share: