Abstract Selective internal radiation therapy (SIRT) with preferential delivery of 90Y microspheres to target lesions has shown promising response rates with limited toxicity in the treatment of hepatocellular (HCC), the second leading cause of cancer death in the world. However, to achieve more durable responses, there is much room to improve/adapt the treatment to ensure that all lesions and lesion sub-regions receive adequate radiation delivery. While externally delivered stereotactic body radiation therapy (SBRT) is well suited for smaller solitary HCC, its application for larger or multifocal disease is challenged by the radiation tolerance of the normal liver parenchyma. A dosimetry guided combined approach that exploits complementary advantages of internal and external radiation delivery can be expected to improve treatment of HCC. To make this transition, however, prospective clinical trials establishing safety are needed. Furthermore, for routine clinic use, accurate and fast voxel-level dose estimation in internal radionuclide therapy, that lags behind external beam therapy dosimetry, is still needed. Our long-term goal is to improve the efficacy of radiation therapy with personalized dosimetry guided treatment. Our objective in this application is to demonstrate that it is possible to use 90Y imaging based absorbed dose estimates after SIRT to safely deliver external radiation to target regions (voxels) that are predicted to be underdosed and to develop deep learning based tools to make voxel-level internal dose estimation practical for routine clinic use. Specifically, in Aim 1, we will perform a Phase 1 clinical trial in HCC patients where we will take the novel approach of using the 90Y PET/CT derived absorbed dose map after SIRT to deliver SBRT to tumor regions predicted to be underdosed based on previously established dose-response models. The primary objective of the trial is to obtain estimates of safety of combined SIRT+SBRT for future Phase II trial design. In parallel, in Aim 2, building on promising initial results we will develop novel deep learning based tools for 90Y PET/CT and SPECT/CT reconstruction, joint reconstruction-segmentation and scatter estimation under the low count-rate setting, typical for 90Y. These methods have a physics/mathematics foundation, where convolutional neural networks (CNNs) are included within the iterative reconstruction process, instead of post-reconstruction denoising. In Aim 3, we will develop a CNN for fast voxel-level dosimetry and combine with the CNNs of Aim 2 to develop an innovative end-to-end framework with unified dosimetry-task based training. At the end of this study, we will be ready to use the new deep learning tools in a Phase II trial to demonstrate enhanced efficacy with SIRT+SBRT compared with SIRT alone and advance towards our long- term goal. This will accelerate adoption of these next-generation tools in clinical practice and will have a significant positive impact because treatment based on patient specific dosimetry will substantially improve efficacy, compared with current standard practice in SIRT. Although we focus on 90Y SIRT, our tools will be applicable in radionuclide therapy in general, a rapidly advancing treatment option.

Narrative We will perform a Phase I clinical trial where standard-of-care Y-90 microsphere radioembolization in hepatocellular carcinoma will be followed by external radiation to target regions that are predicted to be underdosed by Y-90, based on patient specific dosimetry. In parallel, we will develop and test voxel-level internal dosimetry tools using convolution neural networks to make such dosimetry-based planning accurate and fast for routine clinic use. This study is relevant to public health because a dosimetry-guided combination radiation treatment approach is likely to substantially improve patient outcome compared to current standard practice of internal or external radiation only.