MR‐template‐based volumes‐of‐interest generation in individual brain PET of mouse Alzheimer model not only without MR but also without spatial normalization

• Published in: Alzheimer's & dementia Vol. 19; no. S1
• Main Authors: Seo, Seung Yeon, Oh, Jungsu S, Kim, Jae Seung
• Format: Journal Article
• Language: English
• Published: 01.06.2023
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.064020

Background The spatial normalization (SN) of individual mouse brain PET or MRI onto template brain space is one of prerequisites for precise quantitative analysis of positron emission tomography (PET) of mouse brains. This has recently been exploited by deep learning‐based approaches, but is still challenging. Moreover, most of them are based on SN using MR templates, which is not always available in both clinical and preclinical studies. Method In this study, we propose a new only PET‐based deep learning‐based image generation approach to resolve MR‐less PET SN problem. We generated individual‐brain‐specific cortical volumes‐of‐interest (VOI) based on inverse‐spatial‐normalization (iSN) and deep convolutional neural network (deep CNN) models with input of PET images instead of MR images. We applied the proposed methods to the mutated amyloid precursor protein and presenilin‐1 mouse model of Alzheimer’s disease. Eighteen mice underwent T2‐weighted MRI and F‐18 FDG PET scans twice, before and after the administration of human immunoglobulin or antibody‐based treatments. Specifically, skull‐stripped PET images were used as the model inputs and iSN‐based cortical (target) and cerebellar (reference) VOIs were used as the labels. We compared our CNN‐based VOI with conventional (template‐based) VOI by both qualitative (visual) assessment and quantitative correlation analysis of mean cortical SUVR (i.e., cortical uptake normalized by cerebellar uptake) estimated by both methods. Result In qualitative visual assessment, there was no visible difference between CNN‐based VOI and template‐based VOI. The mean count obtained by CNN‐based VOI and template‐based VOI were highly significantly (P<0.001) correlated in target VOI (i.e., cortex). Conclusion Since our deep CNN‐based method successfully generated target VOI, showing concordant results with that of conventional VOI methods in mean count correlation analysis, we believe we have established methods of template‐based VOI not only without spatial normalization but also without MR.