Encrypted login | home

Program Information

Automatic CT-To-CT Contour Segmentation Using Deformable Image Registration Software for Head and Neck (H&N) Cancer Adaptive Radiotherapy

A Kumarasiri

A Kumarasiri*, J Kim, C Liu, F Siddiqui, I Chetty, Henry Ford Health System, Detroit, MI

SU-E-J-100 Sunday 3:00PM - 6:00PM Room: Exhibit Hall

Purpose: The significant amount of time required for target delineation continues to be one of the major challenges associated with adaptive radiotherapy. The aim of this study is to evaluate the accuracy of efficient deformable image registration algorithms (available in the VelocityAI and Varian SmartAdapt commercial systems) for automatic segmentation of physician contours from planning CT to mid-treatment CT images for H&N adaptive radiotherapy.

Methods: Ten head and neck cancer patients were considered for this study, each with a planning CT (CT1) and a second CT (CT2) taken approximately 3 weeks into treatment. Treatment volumes and organs were manually delineated by a physician on both sets of CT scans. B-spline-based VelocityAI and Demons-based SmartAdapt DIR algorithms were used to automatically deform CT1 and the relevant contour sets onto corresponding CT2 images. For each DIR, the volume of interest was set to encompass the whole contour set. The agreement of the automatically propagated contours with manually drawn contours of CT2 was visually evaluated by a physician, and the volume overlap was quantified using DICE coefficients.

Results: The overall mean (1SD) DICE indices were 0.72(0.11) for VelocityAI and 0.68(0.17) for SmartAdapt. Both software attained a high degree of correlation for well differentiated and relatively large organs, with DICE indices often exceeding 0.8. Organs with small volumes and/or those with poorly defined boundaries showed less correlation (DICE: ~0.5), likely due to volume averaging effects. Target volume contours generally aligned well (DICE: 0.7-0.8 for PTVs and 0.8-0.9 for GTVs).

Conclusion: Use of automatic DIR-based contour segmentation in H&N adaptive RT is likely to mitigate the need for manual delineation and thereby improve efficiency. More work is needed to evaluate the accuracy of these tools for routine clinical use, particularly for organs with small volumes, or those with poorly defined boundaries.

Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by Varian Medical Systems, Palo Alto, CA

Contact Email: