Image-based deep-learning prediction of future FDG PET patterns in aging and dementia.

1211 Objectives: This study is to develop an image-based deep-learning technique that automatically generates predicted future positron emission tomography (PET) images based on the patient’s current images and assist the diagnosis of age-related disorders. Methods: Deep-learning architecture was realized in which seven layer U-net convolutional neural network, units normalized by batch normalization, and activated by rectified liner unit. [18F]FDG PET image sets were obtained from 265 subjects ( age 64±8.7 yrs, 93 female) who underwent [18F]FDG PET imaging at multiple time points (total 443 image sets. Additionally, an independent data set (28pair data, 7 subjects, age 64±12.2 yrs, 3 female) were used for testing the accuracy of future image prediction. Both PET slice images and 3D-SSP images were used for the analysis. By learning the relationship between the image at a certain point and the image after one year, this system performs interpolation considering the relation and generates future images. Images obtained in the first year were used to predict images of the subsequent years. In order to compensate for the small amount of data, generative adversarial network (GAN) was used for learning. This study was approved by the institutional review board of Hamamatsu Medical Photonics Foundation (No.86), and written informed consent was obtained from each participant after detailed explanation of the study. Results: The proposed technique achieved root mean square error of 2.5% and peak signal-to-noise ratio of 32.6 dB by using 3D-SSP images when predicting the image at 1-year from the baseline. These results shows that predicted images are similar to the real images. In the baseline and predicted 1-year follow up scans used for training, positive correlation was observed in all cases (frontal right region, frontal left region, occipital right region, and occipital left region) when examining the amount of change in the standardized uptake value (right frontal hemisphere, r=0.41, p<0.001; left frontal hemisphere r=0.39, p<0.001; right occipital hemisphere, r=0.41, p<0.001; left occipital hemisphere; r=0.42, p<0.001; and global brain r=0.41, p<0.001). Similar correlations were observed in the test data (In case of test data (right frontal r=0.35, p<0.05; left frontal r=0.37, p<0.05; right occipital, r=0.34, p<0.05; left occipital, r=0.39, p<0.05; and global r=0.36, p<0.001). Conclusions: This study has demonstrated the feasibility of Image-based deep-learning technique to predict future patterns of [18F]FDG PET scans. This technique could be useful to detect an abnormal course of future changes in the subjects. Such image-based prediction has not been well established in the past. By advancing and improving this method, we believe that it will contribute to future prediction of and early detection of dementia.