RT Journal Article SR Electronic T1 Direct Estimation of Input Function Based on Fine-tuned Deep Learning Method in Dynamic PET Imaging JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 1394 OP 1394 VO 61 IS supplement 1 A1 Liangzhou Wang A1 Tianyu Ma A1 Shulin Yao A1 Qing Ye A1 Jennifer Coughlin A1 Martin Pomper A1 Yong Du A1 Yaqiang Liu YR 2020 UL http://jnm.snmjournals.org/content/61/supplement_1/1394.abstract AB 1394Objectives: Dynamic PET provides more accurate quantitative information and richer disease information than static PET. However, routine clinical application of dynamic PET imaging is limited by invasive arterial blood sampling (or manually annotated image-derived blood activity) for use as an input function. This study aimed to develop a deep learning-based method to directly estimate the input function from dynamic PET data without any manual assistance. Methods: Our study consists of two clinical datasets. The first group of data includes dynamic FDG scans performed on the United Imaging uMI 510 PET/CT system in the Chinese PLA General Hospital. 35 subjects (25 male, 10 females; 13 healthy; age 15-73y) were scanned. Image-derived blood activity was used as input function. The second dataset (n= 26 healthy subjects) includes 90-min dynamic brain data after bolus injection of 11C-DPA-713(DPA) performed on a Siemens HRRT PET system at Johns Hopkins University. All DPA PET data in this study were acquired from individuals with high affinity binding genotype for the 18 kDa translocator protein target and included input function data acquired through arterial blood sampling and radiometabolite measurements. There were 30 dynamic frames in each scan. The input function value for each dynamic frame was interpolated from the measured input function curve. Each dataset was split into 3 parts: 70% as training set, 20% as validation set and 10% as test set. Two deep learning networks, namely the raw model and the fine-tuned model were investigated. The raw model contains a down-sampling convolutional module to extract image features and a fully connected regression module to predict the input function. The 3-D dynamic image and reference image are stacked to form a two-channel 4-D input. For FDG tracer, the reference image refers to image reconstructed from data acquired in 0 to 20 minutes scan. For DPA, the mean image of all dynamic images served as the reference image. L1-loss is chosen as the loss function based on controlled experiment results. In the fine-tuned model, the 1-D vector acquired from the convolutional module are concatenated with medical information including patient’s age, weight, height, injected dose, frame time. Data augmentation methods such as random center crop & resize, rotation, translation and scaling were used in training fine-tuned model. The input function for DPA was normalized according to its integral value before reconstruction. Indirect voxel-based reconstruction methods were used: Patlak analysis with start time (t*) of 20 minutes for Ki (FDG) and Logan analysis with t* of 30 minutes for VT (DPA). To investigate the performance of the proposed method, the root mean square error (RMSE) was calculated between the parametric images reconstructed with true and predicted input function. Results: The predicted input function shows similarity in shape to the standard input function. The prediction error in early stage was larger than in later stage, especially for the peak values in early stage. The fine-tuned model enhances prediction accuracy in the early stage, while the improvement in later stage was not obvious. The visual quality of reconstructed Ki and VT parametric image of the fine-tuned model was apparently better than that of raw model. The RMSE of the fine-tuned model was much lower than the raw model. Conclusions: The proposed fine-tuned deep learning-based method was able to estimate the input function directly from dynamic images. The high accuracy of the reconstructed parametric image using the predicted input function supports pursuit of this method for clinical application. In future work we will investigate to further improve the accuracy of the predicted input function. Support: The research was supported by the National Natural Science Foundation of China (No. 81727807, No.11575096, No. 11605008) and National Key Research and Development (R&D) Plan of China (Grant ID. 2019YFF0302503 and 2016YFC0105405).