Automated CT-CT Deformable Image Registration QA Using Virtual Phantoms
Abstract
Purpose
To address the labor-intensive nature of Deformable Image Registration (DIR) QA quantitatively through the use of independent software and in-house scripting for lung and Head & Neck (H&N) sites for clinical use.
Methods
Ground truth deformations were generated using ImSimQA for 14 Lung (168 deformations) and 7 H&N (231 deformations) patient CTs. Lung deformations included lung volume expansion/shrinkage (low~10mm, medium~20mm, high~30mm) and gross tumor volume expansion/shrinkage (low~5mm, medium~10mm, high~20mm). H&N deformations included head tilt forward/backward/left/right (low~15mm, medium~25mm, high~35mm) and parotid expansion/shrinkage (low~3mm, medium~5mm, high~8mm). To automate data ingestion, a DICOM watcher in MIM was configured to detect and import batch-processed exports from ImSimQA. A custom MIM workflow was implemented to perform registration, export Deformation Vector Fields (DVFs), and automatically calculate Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) between structures deformed by MIM and the corresponding ground truth. This workflow embeds specific keywords into the DICOM series descriptions, triggering downstream Python scripts to automatically identify, pair, organize, and analyze the data. This streamlined process completes the analysis of a single deformation scenario in the order of a few minutes.
Results
The platform successfully processed all contours without manual intervention. For Lung, 1.1 % (315 out of 29105) of the contours tested have 95-percentile DVF larger than 2mm, with 78% of those out of tolerance come from medium or high deformations. For HN, 1.5% (63 out of 4308) of the contours tested have 95-percentile DVF larger than 2mm, with all out of tolerance come from medium or high deformations.
Conclusion
This work establishes a scalable, high-throughput platform for DIR QA using clinical data. By automating data import via DICOM watchers and analysis via keyword-embedded workflows, the labor required for commissioning was significantly reduced. This workflow can be adapted for DIR commissioning of other disease sites with minimum adjustment.