2019 AAPM Annual Meeting
Back to session list

Session Title: Deep Learning for Image Reconstruction and Processing
Question 1: What are the advantages of deep learning reconstruction?
Reference:Lønning, K., Putzky, P. & Welling, M. Recurrent Inference Machines for Accelerated MRI Reconstruction. Int. Conf. Med. Imaging with Deep Learn. 1–11 (2018).
Choice A:Speed of reconstruction compared to analytic methods
Choice B:Possibility for reconstruction of (very) sparse data
Choice C:It is certain the precise details are preserved
Question 2: Which of the following are deep learning models which can directly reconstruct MRI based on the k-space data?
Reference:Lønning, K., Putzky, P. & Welling, M. Recurrent Inference Machines for Accelerated MRI Reconstruction. Int. Conf. Med. Imaging with Deep Learn. 1–11 (2018). Zhu, B., Liu, J. Z., Cauley, S. F., Rosen, B. R. & Rosen, M. S. Image reconstruction by domain-
Choice A:Automap
Choice B:Recurrent inference machines
Choice C:U-nets
Choice D:Sensitivity encoding (SENSE)
Choice E:A and B
Question 3: What is the advantage of deep learning based DBT reconstruction
Reference:Moriakov, N. et al. Deep learning framework for digital breast tomosynthesis reconstruction. in SPIE Medical Imaging (2019).
Choice A:Superior contrast
Choice B:Better in-plane resolution
Choice C:Better coronal resolution
Question 4: Model-enforced post-processing can be advised if, (multiple possible)
Reference:J. Adler and O. Öktem. "Solving ill-posed inverse problems using iterative deep neural networks." Inverse Problems 33.12 (2017): 124007. A. Hauptmann, et al. "Real‐time cardiovascular MR with spatio‐temporal artifact suppression using deep learning�
Choice A:large amounts of training data are available.
Choice B:reconstruction artefacts are uncorrelated/noiselike.
Choice C:quantitative values are important.
Choice D:short computation times are required.
Choice E:superior reconstruction quality is desired.
Choice F:A, B and D.
Question 5: What is the major limitation in high-dimensional iterative reconstructions?
Reference:A. Hauptmann, et al. "Model-based learning for accelerated, limited-view 3-d photoacoustic tomography." IEEE transactions on medical imaging 37.6 (2018): 1382-1393. A. Jonas, and O. Öktem. "Learned primal-dual reconstruction." IEEE transactions on medical imaging 37.6 (2018): 1322-1332.
Choice A:Available memory on the GPU
Choice B:Computational cost of imaging model
Choice C:Acquisition of training data
Question 6: In supervised tomographic image reconstruction, what are suitable loss functions?
Reference:E. Kang, J. Min, and J. C. Ye. "A deep convolutional neural network using directional wavelets for low‐dose X‐ray CT reconstruction." Medical physics 44.10 (2017): e360-e375.
Choice A:Hinge loss
Choice B:Mean squared error
Choice C:Cross entropy
Choice D:Absolute differences
Choice E:Logarithmic loss
Choice F:B and D
Question 7: Which network architecture is especially popular for segmentation in medical imaging?
Reference:O. Ronneberger, P. Fischer, and T. Brox. "U-net: Convolutional networks for biomedical image segmentation." International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015.
Choice A:LeNet
Choice B:AlexNet
Choice C:ResNet
Choice D:U-net
Choice E:YOLO
Question 8: Generative adversarial nets
Reference:I. Goodfellow et al. Generative Adversarial Nets, arXiv 2014
Choice A:are convolutional networks designed for supervised learning
Choice B:are fully connected networks designed for unsuvervised learning
Choice C:are neural networks designed for unsupervised learning
Question 9: Can a neural network be trained to reduce image noise by just receiving pairs of noisy images without seeing a clean (high dose) image
Reference:J. Lehtinen et al. Noise2Noise: Learning Image Restoration without Clean Data, arXiv 2018
Choice A:Yes. The network learning itself is a statistical process and thus a noisy image as input and a matched noisy image as target are sufficient for the training to converge to the ensamble average.
Choice B:Of course not! For supervised learning, matched pairs of low dose and high dose images are necessary to enable the network to learn how to reduce noise. For unsupervised learning, e.g. using GANs, high dose images need to be available to allow the discrimator to learn how low noise images look like. Noisy images as targets are thus useless!
Choice C:No. Otherwise someone would have already applied that to CT images.
Question 10: The real-time calculation of accurate patient dose distributions can be accurately carried out with a 3D U-net
Reference:J. Maier et al. IEEE Medical Imaging Conference Program, M-03-178, November 2018.
Choice A:using the CT volume as the only input to the net
Choice B:using the CT volume and a first order dose estimate as network inputs
Choice C:using the first order dose estimate as the only network input
Back to session list