Improving MRI-based dosimetry for holmium-166 transarterial radioembolization using a nonrigid image registration for voxelwise $\Delta R_2^{\ast }$ calculation

Background: Transarterial radioembolization (TARE) is a treatment modality for liver tumors during which radioactive microspheres are injected into the hepatic arterial system. These microspheres distribute throughout the liver as a result of the blood flow until they are trapped in the arterioles because of their size. Holmium-166 (¹⁶⁶ Ho)-loaded microspheres used for TARE can be visualized and quantified with MRI, as holmium is a paramagnetic metal and locally increases the transverse relaxation rate $R_2^{\ast }$ . The current ¹⁶⁶ Ho quantification method does not take regional differences in baseline $R_2^{\ast }$ values (such as between tumors and healthy tissue) into account, which intrinsically results in a systematic error in the estimated absorbed dose distribution. As this estimated absorbed dose distribution can be used to predict response to treatment of tumors and potential toxicity in healthy tissue, a high accuracy of absorbed dose estimation is required.

Purpose: To evaluate pre-existing differences in $R_2^{\ast }$ distributions between tumor tissue and healthy tissue and assess the feasibility and accuracy of voxelwise subtraction-based $\Delta R_2^{\ast }$ calculation for MRI-based dosimetry of holmium-166 transarterial radioembolization (¹⁶⁶ Ho TARE).

Methods: MRI data obtained in six patients who underwent ¹⁶⁶ Ho TARE of the liver as part of a clinical study was retrospectively evaluated. Pretreatment differences in $R_2^{\ast }$ distributions between tumor tissue and healthy tissue were characterized. Same-day pre- and post-treatment $R_2^{\ast }$ maps were aligned using a deformable registration algorithm and subsequently subtracted to generate voxelwise $\Delta R_2^{\ast }$ maps and resultant absorbed dose maps. Image registration accuracy was quantified using the dice similarity coefficient (DSC), relative overlay (RO), and surface dice (≤4 mm; SDSC). Voxelwise subtraction-based absorbed dose maps were quantitatively (root-mean-square error, RMSE) and visually compared to the current MRI-based mean subtraction method and routinely used SPECT-based dosimetry.

Results: Pretreatment $R_2^{\ast }$ values were lower in tumors than in healthy liver tissue (mean 36.8 s^-1 vs. 55.7 s^-1 , P = 0.004). Image registration improved the mean DSC of 0.83 (range: 0.70-0.88) to 0.95 (range: 0.92-0.97), mean RO of 0.71 (range 0.53-0.78) to 0.90 (range: 0.86-0.94), and mean SDSC ≤4 mm of 0.47 (range: 0.28-0.67) to 0.97 (range: 0.96-0.98). Voxelwise subtraction-based absorbed dose maps yielded a higher tumor-absorbed dose (median increase of 9.0%) and lower healthy liver-absorbed dose (median decrease of 13.8%) compared to the mean subtraction method. Voxelwise subtraction-based absorbed dose maps corresponded better to SPECT-based absorbed dose maps, reflected by a lower RMSE in three of six patients.

Conclusions: Voxelwise subtraction presents a robust alternative method for MRI-based dosimetry of ¹⁶⁶ Ho microspheres that accounts for pre-existing $R_2^{\ast }$ differences, and appears to correspond better with SPECT-based dosimetry compared to the currently implemented mean subtraction method.