Exhibit Hall | Forum 1
Purpose: To quantitatively evaluate the voxel-level uncertainty distribution associated with deformable image registration (DIR) in adaptive radiotherapy.
Methods: We investigated the uncertainty of commercial (Eclipse) and research (LDDMM) DIR algorithms for six lung cancer patients receiving adaptive radiotherapy. For each inquiry patient, six patients from a 60-lung-cancer-patient atlas were semi-automatically matched based on anatomical locations and extent of the GTV. The deformation vector field (DVF) between the atlas planning CT (pCT) and atlas mid-course CBCT was obtained using free-form DIR. Subsequently, we registered the pCTs of atlas and inquiry patient to establish a one-to-one spatial correspondence, and transferred the atlas DVF to the inquiry patient’s corresponding locations. Realistic projection domain noise was also incorporated. Consequently, a digital phantom mimicking a mid-treatment re-simulation CT around the inquiry patient was created with known DVF. Registering phantoms to original pCT allows to compute exact DVF errors and uncertainty at every position. The impact of DVF errors was further evaluated using dice and dose differences of warped contours and planning doses.
Results: The simulated anatomical changes captured the deformations observed between the pCT and re-simulation CT of the actual patients. The mean and standard deviation of the GTV DVF error were 2.3(±1.9)mm and 1.0(±1.3)mm for Eclipse and LDDMM, respectively. Larger errors (>2mm) were observed with Eclipse for locations of low image contrast. For Eclipse, GTV, esophagus, heart and cord dice were 0.93(±0.04), 0.87(±0.05), 0.97(±0.02) and 0.87(±0.05), respectively. LDDM results were significantly better (ANOVA, p<0.05) for all structures. The Eclipse DIR uncertainty demonstrated a larger impact on the esophagus dose (ΔD⁵ᶜᶜ ~1Gy).
Conclusion: The presented data augmentation approach using multiple atlas deformations within the same reference frame allows to capture a large variety of realistic anatomical deformations and enables to calculate the voxel-level DIR error distribution. This can support dose accumulation and plan adaptation decisions.
Registration, Quality Assurance, Deformation