problems – Low-dose x-ray CT (Kang et al, Chen et al, Wolterink et al) – Sparse view CT (Jin et al, Han et al, Adler et al) – Interior tomography (Han et al) – Stationary CT for baggage inspection (Han et al) – CS-MRI (Hammernik et al, Yang et al, Lee et al, Zhu et al) – US imaging (Yoon et al ) – Diffuse optical tomography (Yoo et al) – Elastic tomography (Yoo et al) – etc • Advantages – Very fast reconstruction time – Significantly improved results Other works
matrix comple5on Nuclear norm Projec5on on sampling posi5ons min m kH ( m ) k⇤ subject to P⌦(b) = P⌦( f ) RankH(f) = k * Jin KH et al IEEE TCI, 2016 * Jin KH et al.,IEEE TIP, 2015 * Ye JC et al., IEEE TIT, 2016 m Annihilating filter-based low-rank Hankel matrix
C = T H d(f) C = T (f ~ ) Encoder: ˜ T = I ˜ = PR(V ) H d(f) = U⌃V T Unlifting: f = (˜C) ~ ⌧(˜ ) : Non-local basis : Local basis : Frame condition : Low-rank condition convolution pooling un-pooling convolution : User-defined pooling : Learnable filters H pi (gi) = X k,l [Ci]kl e Bkl i Decoder: Deep Convolutional Framelets (Y, Han, Cha; 2018)
3T / 7T MR imaging dataset. • MRI Data (http://mridata.org/) – Raw k-space dataset acquired on a GE clinical 3T scanner. • LUNA (https://luna16.grand-challenge.org/data/) – Lung Nodule analysis dataset acquired on a CT scanner. • Data Science Bowl (https://www.kaggle.com/c/data-science-bowl-2017) – A thousand low-dose CT images. • NIH Chest X-rays (https://nihcc.app.box.com/v/ChestXray-NIHCC) – X-ray images with disease labels. http://www.cancerimagingarchive.net/ https://www.kaggle.com/datasets • TCIA Collections (http://www.cancerimagingarchive.net/) • De-identifies and hosts a large archive of medical images of cancer accessible for public download. • The data are organized as “Collections”, typically patients related by a common disease, image modality (MRI, CT, etc).
such as enableGpu and enableCudnn. 4. Run the ‘vl_compilenn.m’. * To use GPU processing (false true), you must have CUDA installed. ( https://developer.nvidia.com/cuda-90-download-archive ) ** To use cuDNN library (false true), you must have cuDNN installed. ( https://developer.nvidia.com/cudnn )
need to program the network architecture code because MatConvNet supports the network training framework. Support famous network architectures, such as alexnet, vggnet, resnet, inceptionent, and so on.
layers. • The structure of Stage 0 Layer Name ( string-type ) Layer object ( object ) Input Name ( string-type ) Output Name ( string-type ) Parameters Name ( string-type ) All objects and names must be unique.
default hyper-parameters as follows, Refer the cnn_train.m ( or cnn_train_dag.m ) The supported hyper-parameters 1. The size of mini-batch 2. The number of epochs 3. Learning rate 4. Weight decay factor 5. Solvers such as SGD, AdaDelta, AdaGrad, Adam, and RMS The kind of Optimization Solvers
Jawook Goo US Team • Shujaat Khan • Jaeyong Hur MR Team • Dongwook Lee • Juyoung Lee • Eunju Cha • Byung-hoon Kim Image Analysis Team • Boa Kim • Junyoung Kim Optics Team • Sungjun Lim • Junyoung Kim • Jungsol Kim • Taesung Kwon