• Install conda
• Create conda env $ conda env create -f envs.yml

• At the first datasets must be downloaded, in this work we used ISIC 2018 dataset
• You must download 3 datasets:
  • Training data for tasks 1-2 (10.4 G)
  • Training Ground Truth for task 1 (26 MB) & task 2 (33 MB)
  • Validation Data for tasks 1-2 (228 MB)
  • Validation Ground Truth for task 1 (742 KB) & task 2 (1 MB)
• Create directory named images
• Then unpack these zips into images directory
• Out of the box already works baseline model, to support model with generated images pix2pix generator must be trained first

• go to dataset-to-pix2pix-data folder

• modify the 3rd line of resize-images.sh by filling in the absolute path to the root of this repository

  E.g. REPO_DIR="~/master-diploma" if this repository is located at ~/master-diploma

• execute bash script with arguments:

  $ chmod +x resize-images.sh
  $ DIR=<data-root> ./resize-images.sh -a ISIC2018_Task2_Training_GroundTruth_v3 -s ISIC2018_Task1_Training_GroundTruth -i ISIC2018_Task1-2_Training_Input
  

  where <data-root> is absolute path of the folder images

• go to bounding_boxes folder

• modify the 3rd line of resize-images.sh by filling in the absolute path to the root of this repository

  E.g. REPO_DIR="~/master-diploma" if this repository is located at ~/master-diploma

• execute bash script with arguments:

  $ chmod +x process_images.sh
  $ DIR=<data-root> ./process_images.sh -a ISIC2018_Task2_Training_GroundTruth_v3 -s ISIC2018_Task1_Training_GroundTruth -i ISIC2018_Task1-2_Training_Input
  

  where <data-root> is absolute path of the folder images

• go to the GAN directory $ cd pix2pixHD

• Start training a GAN

  $ python train.py --name <experiment-name> --dataroot <data_root>/datasets/skin --label_nc 8 --checkpoints_dir <directory-to-store-temporary-results> --gpu_id <gpu-id> --batchSize 4
  

  where

  • <experiment-name> is the name by which the trained model will be identified by other scripts
  • <data-root> is absolute path of the folder images
  • <directory-to-store-temporary-results> is name of the directory that will be created by the script under pix2pixHD and where training metadata will be stored
  • is int number - the model will be trained on cuda:<gpu-id>

• If needed, resume training a GAN

  $ python train.py --name <experiment-name> --dataroot <data_root>/datasets/skin --label_nc 8 --checkpoints_dir <directory-to-store-temporary-results> --gpu_id <gpu-id> --batchSize 4 --continue_train
  

All script arguments have the same meaning as in the command above

• After training the GAN, synthesize images

  $ python test.py --name <experiment-name> --dataroot <data_root>/datasets/skin --checkpoints_dir <directory-to-store-temporary-results> --label_nc 8 --how_many 10000 --gpu_id  <gpu-id> --results_dir <data_root>/pix2pix_result/ --phase train
  

  where

  • <experiment-name> is the same as in training the GAN
  • <data-root> is absolute path of the folder images
  • <directory-to-store-temporary-results> is the same as in training the GAN
  • is int number - image synthesis will be performed on cuda:<gpu-id>

• At this step need to create fake images
• go to bboxes folder
• execute python 1_noise_crop.py <path-to-bounding_boxes_metadata.csv> <path-to-folder-images_512p> <base-path-to-storage-fake-images> -- this script will create a lot of fake images
• execute python 2_noise_data_to_pix2pix.py <base-path-to-storage-fake-images> -- process images created on the previous step to create images acceptable by pix2pix
• pass generated data throw pix2pix GAN with command: python3 test.py --name <experiment-name> --dataroot <path-to-lesions-with-masks> --checkpoints_dir <directory-to-storage-temporary-results> --label_nc 8 --how_many 10000 --gpu_id <gpu-id> --results_dir <result-dir>
• execute python 3_create_fake_dataset <base-path-to-storage-fake-images> -- generates csv files
• at the end you'll get set of folders with different strategies to train and execute

• I already split datasets
• use splits folder to train model with usual data
• use splits_boxed folder to train model with bounding boxes
• If you want to create custom splitting with scripts from splits folder

Model based on InceptionV4 network

• go to the classificator directory $ cd classificator_network

• run trainig classifier

  $ python train.py --train_root <data-root-parent> --train_csv <full-path-to-train-csv-image> --validate_root <data-root-parent> --validate_csv <full-path-to-validate-csv-image>  --result_dir <base-result-directory> --experiment_name <launch-name> --epochs 100 --num_workers 0 --batch_size 32 --learning_rate 0.001 --gpu_id <gpu_id>
  

  where

  • <data-root-parent> is absolute path of the folder parent folder of images folder
  • <full-path-to-train-csv-image> is absolute path of the csv with train data
  • <full-path-to-validate-csv-image> is absolute path of the csv with test data
  • <base-result-directory> is relative or absolute path of the folder where results will be stored inside
  • <launch-name> is experiment name
  • <gpu_id> is the id of gpu for training the model

  Note

  1. results will be saved as a json file with metrics, which contains accuracy, f1 measure, AUC values under <base-result-directory>/<launch-name>
  2. final model will be saved under <base-result-directory>/<launch-name>/last_model.pth

  Different running options

  • to run with Bissoto et al.'s train-test split use
    • <full-path-to-train-csv-image>=<repo-root>/splits/baseline_bussio/train_<i>.csv, where <i> is the number of run = 0..9
    • <full-path-to-validate-csv-image>=<repo-root>/splits/validation_skin_lesion.csv
  • to run with original train-test split use
    • <full-path-to-train-csv-image>=<repo-root>/splits/baseline/train_<i>.csv, where <i> is the number of run = 0..9
    • <full-path-to-validate-csv-image>=<repo-root>/splits/validation.csv

  where <repo-root> is absolute path of the root of this repository

• generated data splits creation

  $ cd bounding_boxes
  $ python create_generated_split.py --data-root <data-root> --generated-data-folder <generated-data-folder> --ratios 0.2 0.5 0.8 1.0 --seeds 0 1 2 3 4 5 6 7 8 9
  

  where is absolute path of the images folder is absolute path to the folder that contains generated images