[www] [BibTex]

Note: inproceedings

Abstract: In Magnetic Particle Imaging (MPI), field-free line (FFL) encoding allows for setting up sinograms and use well-known algorithms from Computed Tomography (CT) for image reconstruction. Here, an FFL trajectory with a drive field (DF) direction orthogonal to rotation and translation direction of the FFL is considered. The reconstruction of each DF cycle is mapped to several sinograms along DF direction, which may result in holes within the sinograms. To fill these holes, a CT inpainting approach based on Deep Learning algorithms is adapted. Therefore, different neural network architectures were used. A U-Network, showing good results for inpainting tasks and a generative adversarial network to use a second network for evaluation of image quality. Experiments with different learning rates, architectures, encoders, data augmentation, partial convolution layers and dual domain loss have been performed and evaluated. For training, two data sets were created. From CT data intrathoracic and lower limb vessel structures were segmented to mimic MPI images. Dataset1 presents ideal information, i.e. images were transformed to radon space resulting in ideal sinograms. Dataset2 consists of synthesized MPI measurement data. Each data sets includes 12080 sinograms, split in train (60%), validation (20%) and test (20%) data. Training was started with Dataset1 and one configuration including holes. The optimum was a U-Network with a learning rate of 10-4, early stopping, ResNet50 encoder and partial convolution layers. The pre-trained network was further trained with Dataset2. This improved the performance for actual MPI measurement data. The optimized network was successfully applied for different hole configurations.

[www]

Note: inproceedings

Abstract: In Magnetic Particle Imaging (MPI), field-free line (FFL) encoding allows for setting up sinograms and use well-known algorithms from Computed Tomography (CT) for image reconstruction. Here, an FFL trajectory with a drive field (DF) direction orthogonal to rotation and translation direction of the FFL is considered. The reconstruction of each DF cycle is mapped to several sinograms along DF direction, which may result in holes within the sinograms. To fill these holes, a CT inpainting approach based on Deep Learning algorithms is adapted. Therefore, different neural network architectures were used. A U-Network, showing good results for inpainting tasks and a generative adversarial network to use a second network for evaluation of image quality. Experiments with different learning rates, architectures, encoders, data augmentation, partial convolution layers and dual domain loss have been performed and evaluated. For training, two data sets were created. From CT data intrathoracic and lower limb vessel structures were segmented to mimic MPI images. Dataset1 presents ideal information, i.e. images were transformed to radon space resulting in ideal sinograms. Dataset2 consists of synthesized MPI measurement data. Each data sets includes 12080 sinograms, split in train (60%), validation (20%) and test (20%) data. Training was started with Dataset1 and one configuration including holes. The optimum was a U-Network with a learning rate of 10-4, early stopping, ResNet50 encoder and partial convolution layers. The pre-trained network was further trained with Dataset2. This improved the performance for actual MPI measurement data. The optimized network was successfully applied for different hole configurations.

[www] [BibTex]

Note: inproceedings

Abstract: In Magnetic Particle Imaging (MPI), field-free line (FFL) encoding allows for setting up sinograms and use well-known algorithms from Computed Tomography (CT) for image reconstruction. Here, an FFL trajectory with a drive field (DF) direction orthogonal to rotation and translation direction of the FFL is considered. The reconstruction of each DF cycle is mapped to several sinograms along DF direction, which may result in holes within the sinograms. To fill these holes, a CT inpainting approach based on Deep Learning algorithms is adapted. Therefore, different neural network architectures were used. A U-Network, showing good results for inpainting tasks and a generative adversarial network to use a second network for evaluation of image quality. Experiments with different learning rates, architectures, encoders, data augmentation, partial convolution layers and dual domain loss have been performed and evaluated. For training, two data sets were created. From CT data intrathoracic and lower limb vessel structures were segmented to mimic MPI images. Dataset1 presents ideal information, i.e. images were transformed to radon space resulting in ideal sinograms. Dataset2 consists of synthesized MPI measurement data. Each data sets includes 12080 sinograms, split in train (60%), validation (20%) and test (20%) data. Training was started with Dataset1 and one configuration including holes. The optimum was a U-Network with a learning rate of 10-4, early stopping, ResNet50 encoder and partial convolution layers. The pre-trained network was further trained with Dataset2. This improved the performance for actual MPI measurement data. The optimized network was successfully applied for different hole configurations.

[www]

Note: inproceedings

Abstract: In Magnetic Particle Imaging (MPI), field-free line (FFL) encoding allows for setting up sinograms and use well-known algorithms from Computed Tomography (CT) for image reconstruction. Here, an FFL trajectory with a drive field (DF) direction orthogonal to rotation and translation direction of the FFL is considered. The reconstruction of each DF cycle is mapped to several sinograms along DF direction, which may result in holes within the sinograms. To fill these holes, a CT inpainting approach based on Deep Learning algorithms is adapted. Therefore, different neural network architectures were used. A U-Network, showing good results for inpainting tasks and a generative adversarial network to use a second network for evaluation of image quality. Experiments with different learning rates, architectures, encoders, data augmentation, partial convolution layers and dual domain loss have been performed and evaluated. For training, two data sets were created. From CT data intrathoracic and lower limb vessel structures were segmented to mimic MPI images. Dataset1 presents ideal information, i.e. images were transformed to radon space resulting in ideal sinograms. Dataset2 consists of synthesized MPI measurement data. Each data sets includes 12080 sinograms, split in train (60%), validation (20%) and test (20%) data. Training was started with Dataset1 and one configuration including holes. The optimum was a U-Network with a learning rate of 10-4, early stopping, ResNet50 encoder and partial convolution layers. The pre-trained network was further trained with Dataset2. This improved the performance for actual MPI measurement data. The optimized network was successfully applied for different hole configurations.

[www]

Note: inproceedings

Abstract: In Magnetic Particle Imaging (MPI), field-free line (FFL) encoding allows for setting up sinograms and use well-known algorithms from Computed Tomography (CT) for image reconstruction. Here, an FFL trajectory with a drive field (DF) direction orthogonal to rotation and translation direction of the FFL is considered. The reconstruction of each DF cycle is mapped to several sinograms along DF direction, which may result in holes within the sinograms. To fill these holes, a CT inpainting approach based on Deep Learning algorithms is adapted. Therefore, different neural network architectures were used. A U-Network, showing good results for inpainting tasks and a generative adversarial network to use a second network for evaluation of image quality. Experiments with different learning rates, architectures, encoders, data augmentation, partial convolution layers and dual domain loss have been performed and evaluated. For training, two data sets were created. From CT data intrathoracic and lower limb vessel structures were segmented to mimic MPI images. Dataset1 presents ideal information, i.e. images were transformed to radon space resulting in ideal sinograms. Dataset2 consists of synthesized MPI measurement data. Each data sets includes 12080 sinograms, split in train (60%), validation (20%) and test (20%) data. Training was started with Dataset1 and one configuration including holes. The optimum was a U-Network with a learning rate of 10-4, early stopping, ResNet50 encoder and partial convolution layers. The pre-trained network was further trained with Dataset2. This improved the performance for actual MPI measurement data. The optimized network was successfully applied for different hole configurations.

[www] [BibTex]

Note: inproceedings

Abstract: In Magnetic Particle Imaging (MPI), field-free line (FFL) encoding allows for setting up sinograms and use well-known algorithms from Computed Tomography (CT) for image reconstruction. Here, an FFL trajectory with a drive field (DF) direction orthogonal to rotation and translation direction of the FFL is considered. The reconstruction of each DF cycle is mapped to several sinograms along DF direction, which may result in holes within the sinograms. To fill these holes, a CT inpainting approach based on Deep Learning algorithms is adapted. Therefore, different neural network architectures were used. A U-Network, showing good results for inpainting tasks and a generative adversarial network to use a second network for evaluation of image quality. Experiments with different learning rates, architectures, encoders, data augmentation, partial convolution layers and dual domain loss have been performed and evaluated. For training, two data sets were created. From CT data intrathoracic and lower limb vessel structures were segmented to mimic MPI images. Dataset1 presents ideal information, i.e. images were transformed to radon space resulting in ideal sinograms. Dataset2 consists of synthesized MPI measurement data. Each data sets includes 12080 sinograms, split in train (60%), validation (20%) and test (20%) data. Training was started with Dataset1 and one configuration including holes. The optimum was a U-Network with a learning rate of 10-4, early stopping, ResNet50 encoder and partial convolution layers. The pre-trained network was further trained with Dataset2. This improved the performance for actual MPI measurement data. The optimized network was successfully applied for different hole configurations.