Room: Karl Dean Ballroom B1
Purpose: To improve PET image quality and to enhance the diagnostic accuracy of PET/CT studies by reducing the noise in the reconstructed images through deep learning.
Methods: A novel deep-learning model utilizing Convolutional Neural Networks (CNN) was designed to leverage the information from both the CT and PET images to reduce the noise in PET images. The model includes two sequential blocks of CNN, each containing four convolutional layers with ReLU activation functions. The inputs to the first CNN block were the low-dose PET and the CT image patches of 17x17 pixels. The inputs to the second CNN block were the predicted PET image patches of 9x9 pixels from the first CNN block and the low-dose PET image patches that were fed into the first CNN block. The targets for training were high-dose (4X) PET images. Within each block, the first three convolutional layers contain 64 filters of 3x3 pixels and the last layer contains one 3x3 filter. The neural network was trained and evaluated using phantom and clinical studies.
Results: The structural information in the CT images can aid the definition of the contour of features, and substantially reduce the noise in the low-dose PET images. The neural network can better recover the fine features than both the low-dose and the high-dose PET images. In the clinical studies, the neural networkâ€™s prediction of the signal to noise ratio of the lung lesion was 250% higher than that of the low-dose PET images. The neural network can correctly recover the PET activity despite the strength of the CT signal.
Conclusion: To the best of the authorsâ€™ knowledge, this work is first to demonstrate that deep-learning based noise reduction that employs both the CT and PET information can substantially reduce PET image noise and improve the accuracy of PET images.
Funding Support, Disclosures, and Conflict of Interest: General Electric Company