Supplement Title: Deep learning-based methods for PET image reconstruction and segmentation to enhance radionuclide therapy dosimetry Abstract There is much recent interest in quantitative imaging of yttrium-90 (Y-90) for dosimetry because of the promise of novel Y-90 labelled radionuclide therapies. Deep learning methods are well suited for addressing the challenges of Y-90 positron emission tomography (PET) imaging, where compared with diagnostic FDG PET, true coincidence count-rates are very low while random coincidences are high. The potential of deep learning-based algorithms to outperform conventional algorithms in medical imaging is well recognized, however research in applying these methods to nuclear medicine imaging modalities such as PET is very limited. The few studies applying deep learning to PET imaging have been mostly limited to post-reconstruction image processing/analysis for denoising and feature extraction, and not in the image formation/reconstruction process. Additionally, deep learning research in PET thus far have focused on improving diagnostic imaging, not quantitative imaging, which together with accurate lesion/organ segmentation are pre-requisite for accurate dosimetry. In this supplement, we propose to develop and evaluate deep learning-based image reconstruction and lesion/organ segmentation for low count PET applications such as Y-90 PET. Our approach starts with the raw projection data and utilizes a deep recurrent network in the image formation process. Because the two tasks are mutually dependent, our formalism takes the novel approach of joint reconstruction-segmentation with multi-modality (PET/CT) data. Specifically, we will 1) develop and evaluate Y-90 PET image reconstruction with a deep recurrent network for the regularizer, 2) develop and evaluate deep-learning based joint PET segmentation-reconstruction using multi- modal 90Y PET/CT data. To train/validate/test the proposed methods, we will use clinically realistic phantom measurements and simulations as well as leverage on existing patient data from the parent grant where thus far, PET data and radiologist defined morphologic liver/lesion contours for over 50 cases and 150 lesions are available. We will compare the new Y-90 PET reconstruction with the formulation we recently developed under the parent grant (using conventional untrained regularizers) that showed promising results but suffered from resolution- noise tradeoff. The expected outcome of this work is a well validated deep learning reconstruction- segmentation framework for challenging PET imaging applications where conventional methods are suboptimal.
The proposed research is expected to result in new and accurate tools for quantitative image reconstruction and lesion/organ segmentation that have the potential to contribute significantly toward improving dosimetry in internal radionuclide therapy such as Yttrium- 90 radioembolization in liver malignancies. This study is relevant to public health because a dosimetry-guided personalized approach to radionuclide therapy is highly likely to substantially improve patient outcomes compared to current standard practice. !
