Multi-Center Academic-Industrial Partnership for Personalized Al-Enabled High Count PET

点击翻译按钮获取中文摘要

Abstract High image noise degrades the diagnostic efficacy and quantitative accuracy of PET, as noise could easily results in overestimation of SUV and cause false positive lesion detections in diagnosis. High image noise also decreases the confidence of clinical decision making, leading to additional unnecessary follow-ups through other imaging modalities and invasive procedure. Deep learning-based noise reduction has shown promises for PET imaging. However, existing approaches only focus on converting low-count image (e.g. acquired through low- dose injection or shorter scan time) to standard-count image in typical clinical scans. However, for both low- count and the vast majority of routinely acquired clinical PET images with normal dose and scan time, there is no approach to convert such clinical images to high-count images to further reduce the image noise, mainly due to the challenge of obtaining high-count PET images as training labels. Another challenge in the real-world application is to match the training data with the testing data, in terms of noise level, noise structure, reconstruction parameters, scanner model, etc. Such matching is particularly challenging in a multi-center multi- scanner setting. In this Academic-Industrial Partnership R01 project, we formed an ideal partnership between Visage Imaging, a leading PACS company, and three leading academic centers (Yale, MGH, UC Davis) to develop, evaluate, deploy, and translate robust deep learning methods to generate virtual-high-count PET images in a highly personalized manner by taking into account the noise level of each organ in each patient, as well as associated non-imaging patient information. The academic sites have access to a large number of high- count data that are acquired either through long dynamic scans (at least 90 minutes) or by the ultra-sensitive long axial field-of-view (FOV) Explorer scanner. The developed product would be deep learning networks that can convert any clinical PET images data from all major vendors (Siemens, GE, United Imaging Healthcare (UIH)) into virtual-high-count ultra-low noise images. Since Yale, MGH, and UC Davis are all serviced by Visage Imaging, the developed deep learning technique can be seamlessly translated into Visage PACS research/clinical servers for validation and evaluation, beta testing and user feedback, and ultimate translation and regulatory filings. In Aim 1, we will develop deep learning models for virtual-high-count PET generation. In Aim 2, we will evaluate and deploy the models into Visage research PACS server and evaluate virtual-high-count PET in clinical environments. In Aim 3, we will integrate the developed virtual-high-count PET deep learning models into the clinical production PACS server and generate regulatory documents and supporting data for FDA 510(k).