Y-90 Bremsstrahlung SPECT Scatter Estimation and Voxel Dosimetry Using a Deep Learning Pipeline

Introduction: Y-90 SPECT based dosimetry post-liver radioembolization (RE) is challenging due to the inherent scatter and the poor spatial resolution of Bremsstrahlung SPECT. Monte Carlo (MC) radiation transport algorithms are recognized as the gold standard for both scatter estimation and voxel dosimetry but are computationally expensive and limited by the SPECT image quality. Our goal was to explore a deep-learning based absorbed dose-rate estimation method for Y-90 that can address the fundamental accuracy-efficiency trade-off inherent in current scatter estimation and voxel dosimetry approaches.

Methods: This study presents a pipeline consisting of three distinct stages for Y-90 SPECT-based dosimetry. 1) our previous convolutional neural network-based (CNN) scatter estimation [Eur J Nuc Med 2020;47:2956-67], 2) in-house (OS-EM) algorithm for SPECT reconstruction with the CNN scatter estimate for scatter correction (SC), 3) extending our deep residual learning network, DblurDoseNet [Med Phys 2022;49:1216-30] to Y-90 to estimate the absorbed dose-rate map while compensating for the limited resolution of SPECT. For training and testing the pipeline, 18 (12 for training and 6 for testing) digital phantoms (virtual patients) were generated using clinical Y-90 SPECT/CT data acquired post-RE. In the first stage, we generated ground-truth (GT) scatter projections for training and validation, by running the SIMIND MC simulation code that couples the digital phantoms with the SPECT/CT camera model. The network uses the projected CT-based attenuation map and the total SPECT projections to predict the scatter projections. The second stage reconstructs the SPECT data, with the scatter estimates coming from the first network. For the final stage, the GT dose-rate maps were obtained by running our in-house MC dosimetry code (Dose Planning Method, DPM) with the true phantom activity/density maps as the input. Next, we input the activity map from SPECT reconstruction and the CT-based density map, as well as an initial dose-rate map from Dose Voxel Kernel (DVK) convolution (a more efficient alternative to MC but is suboptimal in the presence of tissue heterogeneities) into the DblurDoseNet and trained it to learn the difference between the GT and DVK dose-rate maps (residual learning). We evaluated the CNN for scatter estimation in both the projection and image domain by Normalized Root Mean Squared Error (NRMSE). Then, we evaluated the complete pipeline (CNN SC + DblurDoseNet) using the normalized Mean Absolute Error (MAE) and NRMSE of dose-rate in lesions/livers. We compared the performance of our pipeline to traditional voxel-dosimetry methods.

Results: When testing on virtual patient phantoms, the NRMSE of CNN-estimated scatter projections was 4.0%±0.7%. SPECT reconstruction with CNN SC had a NRMSE of 10.0%±1.7% vs 149.5%±32.6% for SPECT with no SC typically used in the clinic due to lack of efficient methods for Bremsstrahlung SC. When testing the combined pipeline, CNN SC + DblurDoseNet achieved a %MAE of 7.7%±5.7% and 6.1%±5.0% for dose-rate in lesions and healthy livers respectively, while the corresponding results for CNN SC + DVK/MC were 24.0%±5.9% / 24.0%±5.9% and 15.2%±6.1% / 15.1%±6.5%. For NRMSE of dose-rate in lesions and healthy livers respectively, CNN SC + DblurDoseNet achieved 20.1%±4.8% and 27.9%±3.9%, whereas CNN SC + DVK/MC achieved 27.6%±5.4% / 27.6%±5.4% and 29.7%±4.0% / 29.7%±4.0%. Overall, the improvement in results for CNN SC + DblurDoseNet over CNN SC + DVK/MC was 65.9% / 65.8% for MAE and 19.8% / 19.6% for NRMSE.

Conclusions: Our combined deep learning pipeline for Y-90 SPECT generated more accurate dose-rate maps (achieving %MAE of < 8% on average in lesions and liver) than traditional (non-learning) methods, by learning the scatter and compensating for the poor spatial resolution of bremsstrahlung SPECT.

• Download figure
• Open in new tab
• Download powerpoint