Statistical modeling approach to quantitative analysis of interobserver variability in breast contouring

Area of Science:

• Medical Physics
• Radiation Oncology
• Biostatistics

Background:

• Accurate delineation of organs at risk and target volumes is crucial for effective radiation therapy.
• Interobserver variability in contouring can significantly impact treatment planning and patient outcomes.
• Standardized contouring practices are essential for reproducible and high-quality cancer care.

Purpose of the Study:

• To develop and validate a novel statistical approach for quantifying interobserver variability in breast cancer radiation therapy contouring.
• To assess the impact of contouring from scratch versus using a deformable image registration-generated template on variability.
• To identify individual radiation oncologists whose contouring practices deviate from the consensus.

Main Methods:

• Eight radiation oncologists contoured left breast tumors, both from scratch and using a template derived from deformable image registration.
• The simultaneous truth and performance level estimation algorithm was used to generate a group consensus contour.
• Individual Jaccard indices were fitted to a beta distribution model to quantify variability.
• The analysis was extended to include 9 oncologists from 8 institutions contouring 2 additional patients.

Main Results:

• Contouring from scratch showed broad interobserver variability (mean Jaccard index 86.2%, SD ±5.9%).
• Radiation Therapy Oncology Group-trained physicians demonstrated higher agreement with the group consensus.
• Contouring using a template significantly reduced variability (mean Jaccard index 92.3%, SD ±3.4%).
• The statistical model identified individual oncologists with consistently larger or smaller contours compared to the group.

Conclusions:

• The proposed statistical method effectively quantifies interobserver variability in radiation therapy contouring.
• This approach provides valuable feedback to individual oncologists, enabling them to adjust their practices for improved consistency.
• Utilizing templates generated by deformable image registration can substantially enhance contouring agreement among physicians.