%0 Journal Article %A Fei Gao %A Sven Zuehlsdorff %T An AI system using hybrid CNN and Radiomics model to improve PET image quality %D 2019 %J Journal of Nuclear Medicine %P 1216-1216 %V 60 %N supplement 1 %X 1216Objectives: PET/CT studies typically involve complex, configurable acquisition and reconstruction workflows. A deep Convolution Neural Network (CNN) based Artificial Intelligence (AI) system has previously demonstrated the potential for improving the efficiency of a traditional workflow by automating steps that often require user interaction [1]. However, various clinical scenarios, e.g. rare cases, image artifacts, user errors etc. may compromise the accuracy of the AI system; therefore, increasing accuracy is always desired. The assessment of Radiomics features from the image alone is considered inferior to CNN; however, the combination of quantitative hand-crafted Radiomics features and deep learning potentially further improves CNN based medical imaging analysis by adding information that characterizes the data [2]. Following previous work in [1], the objective of this study is to demonstrate proof of concept by using a hybrid CNN and Radiomics model in an AI system to further improve the accuracy of automatically determining advanced scatter correction parameters to optimize image quality for clinical reading. Methods: Anonymized 18F-FDG and 68Ga-PSMA PET images of a total of 83 patients from multiple institutions were used. A total of 729 Region of Interests (ROIs) have been identified based on the factors sensitive to scatter estimation (contrast, signal to noise ratio etc.) to train a hybrid model. Figure 1 illustrates the model design, which combines image features from deep CNN model as in [1] and 107 hand-crafted Radiomics features (PyRadiomics, [3]). These Radiomics features have been selected by order first order, glcm, gldm, glrlm, glszm, shape features, and represent the specific characteristics of ROIs to differentiate marginal cases. Subsequently, the trained model was used as in [1] to perform an image quality check step and potentially improve the image quality. Results: The available data was split into a training data set and testing data set (80%/20%, respectively). With the same setup in the CNN models, the accuracy increased from 93.8% in single CNN model to 96.4% in the hybrid model to recommend scatter correction parameters that resulting in best image quality for a given case. The area under the ROC curve (AUC) increased from 0.972 in single CNN model to 0.990 in the hybrid model. Conclusions: With added data-characterization Radiomics features, the proposed AI system using hybrid CNN and Radiomics model has the potential for further improving the accuracy of a traditional single CNN based model especially when with limited datasets. The design of the AI model can be expanded to further improve workflow and image quality. Research Support: None [1] Gao, Fei, et al. "An AI system to determine reconstruction parameters and improve PET image quality." Journal of Nuclear Medicine 59.supplement 1 (2018): 31-31. [2] Li, Zeju, et al. "Deep learning based radiomics (DLR) and its usage in noninvasive IDH1 prediction for low grade glioma." Scientific reports 7.1 (2017): 5467. [3] van Griethuysen, Joost JM, et al. "Computational radiomics system to decode the radiographic phenotype." Cancer research 77.21 (2017): e104-e107. %U