Style transfer architecture for inferring pixelwise elastic modulus values of human tissue samples from grayscale microscope images. Based on the Pix2Pix architecture implementation in Keras by Erik Linder-Norén.

• train.py: takes the following command line arguments:

  • --dataset_name - the dataset to train on (default: 'liver')
  • --epochs - integer number of training epochs (default 10000)
  • --batch_size - integer batch size (default 16)
  • --model_name - (required) the .h5 weights file to save to (without the file extension)

• predict.py: takes the following command line arguments:

  • --fn - (required) the filename of the unstained TIFF image to predict on (without the file extension)
  • --nr and --nc - the integer number of subrows and subcolumns to split the image into to save memory (default 8)
  • --model_name - (required) the saved .h5 weights file to use for prediction (also without the file extension)

• data_loader.py: DataLoader class is used during training and testing for loading batches of images, preprocessing them and performing data augmentation as well as processing subimages during prediction.

• GAN.py: instantiates the GAN class. Edit this to change the network architecture.

Tested on Windows 10 and Ubuntu 18.04.3 LTS.

To fully test, follow all the below steps. To test prediction only using a pre-trained model on a test image (both found in the testing folder), skip steps 3 and 4. Installation should take less than 20 minutes on a normal desktop computer with CUDA already installed (see here). Note that a normal tensorflow installation can also be used but training and prediction will take significantly longer

1. Install Python 3.5.2/3.68 with the following packages:
   • tensorflow-gpu (tested with version 1.11.0/1.12.0 using CUDA 9.0)
   • tensorflow (only if your computer does not have a NVIDIA GPU - do not install tensorflow-gpu as well)
   • keras (tested with version 2.24)
   • h5py
   • Pillow
   • numpy
   • opencv-python
2. Clone the github repo
3. Download the training dataset here and extract into the cloned folder
4. Run
   • train.py --dataset_name liver --epochs 10000 --batch_size 16 --model_name liver
5. Once it has finished training (unless training was skipped), move test.tiff into the main directory and run
   • predict.py --fn test --nr 8 --nc 8 --model_name liver

The output will be a file named test_pred.tiff, however it is a 32-bit float image so will not display properly in most photo editors. test_pred.png is provided in the testing folder as a prediction TIFF that been loaded in ImageJ, converted to a pseudocolor image using a lookup table and adjusted so that only values between 0 and 2 kPa are displayed. The image mask used to remove pixels corresponding to bubbles, tissue glue, background or out of focus regions is provided in test_mask.tiff which is also provided as a PNG file. The masked prediction test image is provided as test_pred_masked.png.

Typical times to train on the liver dataset:

• 3 hours on NVIDIA Tesla V100-DGXS 32GB (Ubuntu 18.04)
• 5 hours on NVIDIA GeForce GTX 1050 Ti 4GB (Windows 10)
• 19 hours on CPU (Intel i9-8950HK CPU, Windows 10)

Typical times to predict on test.tiff (3672 x 3282 11.4 MB image):

• 14 minutes on NVIDIA Tesla V100-DGXS 32GB (Ubuntu 18.04)
• 18 minutes on NVIDIA GeForce GTX 1050 Ti 4GB (Windows 10)
• 41 minutes on CPU (Intel i9-8950HK CPU, Windows 10)

To cite this code, please use the following: