This is a reference implementation for paper "Synthesizing 7T from 3T MRI via Deep Learning in Spatial and Wavelet Domains" (caffe implementation). In this paper, we introduce a novel deep learning network that fuses complementary information from spatial and wavelet domains to synthesize 7T T1-weighted images from their 3T counterparts.
The code has been tested on Linux (Ubuntu 14.04), using Caffe, Titan X GPU and on MATLAB 2018b.
Code is provided by Liangqiong Qu on 2019/08/16)
-
Caffe and matcaffe: Change the makefile.config file according to the your own computer and build caffe as "http://caffe.berkeleyvision.org/installation.html"
-
built code as: "make all -j8 && make matcaffe"
- Pre-processing of the dataset (including min-max normalization, histogram matching, and data augmentation)
Run WATNet/matlab/demo/demo_processing_hdf5_for_train.m - Note that this demo also contains instruction for preparing hdf5 file. Step 3 is not required when testing
- Step 1: Preparing hdf5 file for Data processing
Run WATNet/matlab/demo/demo_processing_hdf5_for_train.m - Step 2: cd to folder WATNet/examples/MRI/
% please change the files in AF_encod_3layer_decod_3input.prototxt to your own prepared hdf5 files
Run bash train_MRI.sh
- Important in step 2: Make sure changing the root in AF_encod_3layer_decod_3input.prototxt to your own prepared hdf5 root
Run WATNet/matlab/demo/test_AF_encod_3layer_decod_3input.m
- Pre-trained models and protoxt files for model definition are saved in folder WATNet/examples/MRI/
- WATNet was trained with the image patches extracted from 3T and 7T images. The patch size was 64 × 64 × 3 (or lager size for better performance), covering 3 consecutive axial slices to promote inter-slice continuity
- Train with crop size 64×64×3, and then finetune on size 128×128×3 or even more lager size help produce better performance
- Test code are saved in folder "WATNet/matlab/demo"
- Initial results are saved in folder "WATNet/matlab/demo/Final_result/", you can use these results for comparisons