Lung ultrasound images synthesis using DCGAN
This code will train for 1000 epochs a DCGAN from images provided by a specified folder/directory (TRAIN_IMAGES_DIRECTORY) and save a certain number of fake images (AMOUNT) wherever the user needs (OUTPUT_FAKE_IMAGES_DIRECTORY).
(Note: It will also generate a CSV file with the loss and accuracy results.)
Execution (WSL):
python3 GAN_train.py TRAIN_IMAGES_DIRECTORY OUTPUT_FAKE_IMAGES_DIRECTORY AMOUNT
Example
python3 GAN_train.py DB_LUS4MELUS/tif/normal generated_test 3000
The training will also save the trained generator model (as a h5 file) if the user wants to use it in order to generate more images in the future. (See "Load and use trained generator" below). The file will be named after the folder specified when trained. Following the last example, the generator file will be named as:
generator_generated_test.h5
If wanted, the user can generate more "fake" images with the generator trained on his last run.
Execution (WSL):
python3 load_generator.py GEN_FILE_NAME AMOUNT
Example
python3 load_generator.py generator_generated_test.h5 1000
This code will calculate the FID and KID between 2 set of images, so you can compare the generated images within FAKE_IMAGES_DIRECTORY against the real ones in REAL_IMAGES_DIRECTORY.
(Note: The resulting FID/KID numbers will be displayed on the terminal)
Execution (WSL):
python3 image_evaluation.py REAL_IMAGES_DIRECTORY FAKE_IMAGES_DIRECTORY
Example
python3 image_evaluation.py DB_LUS4MELUS/tif/normal generated_test
This code will just save two PNG images of the Loss and Accuracy plot using the CSV generated on the training process.
Execution (WSL):
python3 results_plot.py
- It is strongly recommended that the directories used for evaluation only contain the images necessary.
- KID numbers make more sense comparing an equal amount of samples for each group.
- Made in Tensorflow 2.1.0 (Keras 2.2.4-tf)
- Requirements and hardware (WIP)