@article{DawoodBreuerStebanietal.2023,
  author    = {Dawood, Peter and Breuer, Felix and Stebani, Jannik and Burd, Paul and Homolya, Istv{\´a}n and Oberberger, Johannes and Jakob, Peter M. and Blaimer, Martin},
  title     = {Iterative training of robust k-space interpolation networks for improved image reconstruction with limited scan specific training samples},
  series = {Magnetic Resonance in Medicine},
  volume    = {89},
  journal   = {Magnetic Resonance in Medicine},
  number    = {2},
  doi       = {10.1002/mrm.29482},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-312306},
  pages     = {812 -- 827},
  year      = {2023},
  abstract  = {To evaluate an iterative learning approach for enhanced performance of robust artificial-neural-networks for k-space interpolation (RAKI), when only a limited amount of training data (auto-calibration signals [ACS]) are available for accelerated standard 2D imaging. Methods In a first step, the RAKI model was tailored for the case of limited training data amount. In the iterative learning approach (termed iterative RAKI [iRAKI]), the tailored RAKI model is initially trained using original and augmented ACS obtained from a linear parallel imaging reconstruction. Subsequently, the RAKI convolution filters are refined iteratively using original and augmented ACS extracted from the previous RAKI reconstruction. Evaluation was carried out on 200 retrospectively undersampled in vivo datasets from the fastMRI neuro database with different contrast settings. Results For limited training data (18 and 22 ACS lines for R = 4 and R = 5, respectively), iRAKI outperforms standard RAKI by reducing residual artifacts and yields better noise suppression when compared to standard parallel imaging, underlined by quantitative reconstruction quality metrics. Additionally, iRAKI shows better performance than both GRAPPA and standard RAKI in case of pre-scan calibration with varying contrast between training- and undersampled data. Conclusion RAKI benefits from the iterative learning approach, which preserves the noise suppression feature, but requires less original training data for the accurate reconstruction of standard 2D images thereby improving net acceleration.},
  language  = {en}
}
@article{AnkenbrandLohrSchloetelburgetal.2021,
  author    = {Ankenbrand, Markus Johannes and Lohr, David and Schl{\"o}telburg, Wiebke and Reiter, Theresa and Wech, Tobias and Schreiber, Laura Maria},
  title     = {Deep learning-based cardiac cine segmentation: Transfer learning application to 7T ultrahigh-field MRI},
  series = {Magnetic Resonance in Medicine},
  volume    = {86},
  journal   = {Magnetic Resonance in Medicine},
  number    = {4},
  doi       = {10.1002/mrm.28822},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-257604},
  pages     = {2179-2191},
  year      = {2021},
  abstract  = {Purpose Artificial neural networks show promising performance in automatic segmentation of cardiac MRI. However, training requires large amounts of annotated data and generalization to different vendors, field strengths, sequence parameters, and pathologies is limited. Transfer learning addresses this challenge, but specific recommendations regarding type and amount of data required is lacking. In this study, we assess data requirements for transfer learning to experimental cardiac MRI at 7T where the segmentation task can be challenging. In addition, we provide guidelines, tools, and annotated data to enable transfer learning approaches by other researchers and clinicians. Methods A publicly available segmentation model was used to annotate a publicly available data set. This labeled data set was subsequently used to train a neural network for segmentation of left ventricle and myocardium in cardiac cine MRI. The network is used as starting point for transfer learning to 7T cine data of healthy volunteers (n = 22; 7873 images) by updating the pre-trained weights. Structured and random data subsets of different sizes were used to systematically assess data requirements for successful transfer learning. Results Inconsistencies in the publically available data set were corrected, labels created, and a neural network trained. On 7T cardiac cine images the model pre-trained on public imaging data, acquired at 1.5T and 3T, achieved DICE\(_{LV}\) = 0.835 and DICE\(_{MY}\) = 0.670. Transfer learning using 7T cine data and ImageNet weight initialization improved model performance to DICE\(_{LV}\) = 0.900 and DICE\(_{MY}\) = 0.791. Using only end-systolic and end-diastolic images reduced training data by 90\%, with no negative impact on segmentation performance (DICE\(_{LV}\) = 0.908, DICE\(_{MY}\) = 0.805). Conclusions This work demonstrates and quantifies the benefits of transfer learning for cardiac cine image segmentation. We provide practical guidelines for researchers planning transfer learning projects in cardiac MRI and make data, models, and code publicly available.},
  language  = {en}
}