Evaluation of a deformable image registration algorithm for image‐guided thermal ablation of liver tumors on clinically acquired MR‐temperature maps

Abstract

Abstract Background Quantitative real‐time MRI‐based temperature mapping techniques are hampered by abdominal motion. Intrascan motion can be reduced by rapid acquisition sequences such as 2D echo planar imaging (EPI), and inter‐scan organ displacement can be compensated by image processing such as optical flow (OF) algorithms. However, motion field estimation can be seriously affected by local variation of signal intensity on magnitude images inherent to tissue heating, potentially leading to erroneous temperature estimates. Purpose This study aims to characterize, in the context of clinical MRI‐guided microwave ablation (MWA), a novel deformable image registration (DIR) algorithm that enhances the generation of thermal maps aligned to a reference position, a critical step for calculating cumulative thermal dose and, consequently, for the real‐time evaluation of interventional procedure progress. Methods A retrospective image analysis was performed on 11 patients that underwent MWA of a liver tumor (primary or metastasis). Ablation duration was set to 9 ± 2 min with a 14‐gauge large antenna. A stack of 13–20 contiguous slices was acquired dynamically (350 repetitions) at 1.5T using a single‐shot EPI sequence. Evaluation was first performed on motion‐free datasets (5 gated acquisitions using a cushion positioned in the patient abdomen) then with ones with motion (8 fixed‐frequency acquisitions at 0.5 Hz). Temperature, thermal dose and lesion size were computed using three workflows: (i) standard phase subtraction (gold standard), (ii) conventional OF motion compensation, (iii) PCA‐based OF motion compensation. The impact of flow field, temperature and lesion volume estimation were compared using averaged endpoint error (AEE), NRMSE and bland Altman plot, respectively. Results Intensity signal decreases (close to 50%) were observed in the vicinity of the probe during MW energy delivery. Both motion correction algorithms reduce the NRMSE of magnitude images throughout the acquisition ( p < 0.005) and achieve similar results between them. Gated acquisition results . Conventional OF produced erroneous vector fields compared to the PCA‐based OF, leading to higher maximal EE (3 mm vs. 1 mm) and temperature errors up to 15°C–20°C. PCA‐based OF algorithm significantly reduces the NRMSE of temperature ( p < 0.005). The conventional OF method underestimated the final size of lesions with a bias of 0.93 cm 3 while the PCA‐based OF reported a bias of 0.5 cm 3 . Fixed frequency acquisition results . The temperature estimation without motion correction led to strong fluctuations or loss of temperature measurement while the proposed PCA‐based OF recovered both a stable and precise measurement with null bias. Conclusion The deformable image registration algorithm is less sensitive to local variations of the signal. Volumetric temperature imaging without gating (20 slices/2 s) could be performed with the same accuracy, and offer trade‐offs in acquisition time or volume coverage. Such a strategy is expected to increase procedure safety by monitoring large volumes more rapidly for MR‐guided thermotherapy on mobile organs.