RT Journal Article
SR Electronic
T1 Strategy to acquire high resolution PET images with Super-Resolution Convolutional Neural Network (SRCNN)
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 105
OP 105
VO 61
IS supplement 1
A1 Chietsugu Katoh
A1 Endo Daiki
A1 Keiichi Magota
A1 Osamu Manabe
A1 Kenji Hirata
YR 2020
UL http://jnm.snmjournals.org/content/61/supplement_1/105.abstract
AB 105Objectives: Super-Resolution Convolutional Neural Network (SRCNN) is a technique for generating a high-resolution image from a low-resolution image using deep learning algorithm. A high-resolution PET device was introduced, high-definition PET data of 2 mm pixels can be obtained in addition to conventional 4 mm pixel PET data. At present, the mainstream PET devices for clinical use are those that output images of 4 mm pixels. Therefore, we performed that a high-definition PET image of 2mm pixels could be generated from a general PET device using SRCNN technology. Methods: Fifteen subjects with malignant tumors were employed. They all had 18F-FDG whole body PET study using a semiconductor PET/CT which yielded 2 mm pixel high-resolution PET data in addition to the conventional 4 mm pixel PET data. Low-resolution short-axis images of 144×144 pixels (one voxel size 4×4×4mm) (250 images per subject) and high-resolution short-axis images of 288 × 288 pixels (one voxel size 2×2×2mm) (500 images per subject) were obtained. Deep learning was performed using low-resolution images as input data, and high-resolution images were applied as teacher data. By changing the number of teacher data, epoch number and mean squared error (MSE) were evaluated. For the SRCNN analysis, a computer was used with following devices; CPU: Intel Core i9 9900K, Memory 64GB, GPU: GeForce 2080Ti (11GB). The development environments were as follows; OS: Windows10, language: Python 3.7, framework: Keras 2.2.4, TensorFlow 1.14.0. Evaluation of quantification of the SRCNN was performed with NEMA (National Electrical Manufactures Association) tumor phantom using SRCNN model trained with 10 subjects’ teacher data and 5 subjects’ verification data. Six spherical phantoms with a diameter of 10 to 37 mm were filled with 18F, whose standardized uptake value (SUV) were set to be eight. Results: SRCNN model was trained with 10 subjects’ teacher data and 5 subjects’ verification data. The trained loss curve showed low MSE value enough to estimate high-resolution images. MSE of 50 randomly extracted 4mm pixel images and the corresponding 2mm pixel image was 1.54±1.31, MSE of 2mm pixel images from SRCNN and the corresponding 2mm pixel images was 0.40±0.31. The latter showed a significantly lower value than the former (p <0.001). A high-resolution 2mm pixel image from SRCNN of NEMA tumor phantom showed better partial volume effect than 4mm pixel image. Conclusions: SRCNN dedicated high-resolution PET images using conventional low-resolution PET images. This technique would be useful for clinical PET study.