RT Journal Article SR Electronic T1 Machine learning based quantitative analysis: an alternative method for traditional kinetic modeling based quantitative analysis of dynamic PET data JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 407 OP 407 VO 60 IS supplement 1 A1 Leyun Pan A1 Xun Zhang A1 Panagiota Koralli A1 Caixia Cheng A1 Eric Ke Wang A1 Yunming Ye A1 Antonia Dimitrakopoulou-Strauss YR 2019 UL http://jnm.snmjournals.org/content/60/supplement_1/407.abstract AB 407Objectives: Traditionally, iterative fitting based kinetic modeling method is the option commonly used for quantitative analysis of dynamic PET data . Essentially, the kinetic modeling is a processing of dimensionality reduction. The lower-dimension kinetic parameters are calculated from high-dimension time activity data (TAD) and are then used for quantitative analysis like correlation, regression and classification, which is a so called two-phase analysis method. One of the reasons for using kinetic modeling is the limited computation ability in former years. From the machine-learning point of view, the original high-dimension TAD are more suitable than the model parameters dataset. With the emerging of more powerful computation systems and the new machine learning methods, direct quantitative analysis from original TAD is becoming possible. Methods: We built a kinetic modeling database for reference that consists of 2701 VOI studies in oncological patients evaluated by kinetic modeling experts. The database consists of 3 parts: Dataset1 is the input (TAD in blood); Dataset2 is the target (TAD in tumor tissue or normal tissue) and Dataset3 is the kinetic modeling parameters that are calculated from Dataset1 and Dataset2 by kinetic modeling method. Each tissue is tagged with the category 0/1(normal/tumor). Purpose of the study was to compare the classification capability based on two methods. Method 1 is the direct machine learning based classification (DMLC) method using Dataset1+Dataset2. Method 2 is the two-phase kinetic modeling based classification (2KMC) method using Dataset3. Recurrent neural network (RNN) is used as the classification tool for both methods in the study. Considering the time relations in TAD dataset, Long Short Term Memory network (LSTM) is used for learning long-term dependencies. After enhancing the ability of generalization by using dropout layer, we added a logistic regression layer to the end of our model for binary classification. Furthermore, suitable RNN parameters of the model are obtained from several adjustments and tunings. Results: The results [TABLE] showed that the DMLC method has a relative better performance. For example, the accuracy of classification based on the DMLC method is 0.8746 compared to 0.8675 for the 2KMC method. Even without input Dataset 1, target Dataset 2 demonstrates a better classification capability with an accuracy of 0.8722. Conclusions: Direct machine learning based classification (DMLC) method using the original TAD data achieved a relative better accuracy as compared with the two-phase kinetic modeling classification (2KMC) method. The results demonstrate that the kinetic modeling is not necessary for quantitative analysis of dynamic PET data with the new machine learning based method. Without the complicated and user-dependent kinetic modeling, the DMLC method is a more robust and reproducible solution. Next step is to apply the method for pixel-based parametric imaging to produce 0/1(normal/tumor) mask parametric images. View this table:Classification performance of DMLC method and 2KMC method