Protocol for performing deep learning-based fundus fluorescein angiography image analysis with classification and segmentation tasks
- Ubuntu 18.04 or higher, Windows 10 or higher with Anaconda or Miniconda
- Python 3.7
- PyTorch = 1.11.0
- tensorflow-gpu = 1.14.0
- CUDA 11.3
- More packages please refer to requirement.txt
a. Create a conda virtual environment and activate it, e.g.,
conda create --name [your_name] python=3.7
conda activate [your_name]
b. Clone the FFA repository.
git clone https://github.com/huapu4/FFA.git
cd FFA
c. Install the environments.
pip install -r requirement.txtThere will be three datasets to be prepared, respectively in folder: 01.Phase_identification , 02.Disease_diagnosis 03.Area_segmentation.
a. Put your own data in ./origin_data, stored according to specifications.
cd origin_data
tree -d -L 2
.
├── 01.phase_identification
│ ├── arterial_phase
│ ├── non_ffa
│ └── venous_phase
├── 02.disease_diagnosis
│ ├── brvo
│ ├── none_np
│ ├── normal
│ └── with_np
└── 03.area_segmentation
├── labeled_data
└── voc_data
b. For 01.Phase_identification and 02.Disease_diagnosis, they are classification task, with the same data deployment.
cd FFA
python cls_data_deploy.py --task [name_of_task] --input [origin_data_folder] --output [output_folder] --prop [proportion of trainset]
c. For 03.Area_segmentation, it is a segmentation task, please run.
python seg_data_deploy.py --input [VOC_data_folder] --output [output_folder] --prop [proportion of trainset]
a. before training, augments and hyperparameters in [config.py] as follows:
--train_data_root Train set path.
--test_data_root Test set path.
--exter_data_root external test set path (if has).
--batch_size the learning amount of each batch.
--test_batch_size the testing amount of each batch.
--num_workers how many workers for loading data.
--resize_value set the image icls_3ut size.
--train_log the path of train log in tensorboard record.
--test_log the path of test log in tensorboard record.
--max_epoch the maximun epoch in this task.
--lr = 0.001 learning rate.
# if you have mutiple gpu devices, so you may set under two param.
--device_ids select which devices， e.g. gpu0,gpu1,gpu4->[0,1,4].
--pri_device select the primary GPU number.
b. begin training
cd 01.Phase_identification
python train.py
#----------- or -----------
cd 02.Disease_diagnosis
python train.py
c. validation
python validation.py --dataset [path of validset] --model [path of *.pth] --excel [name of excel] --roc [name of roc_image]
a. parameters are as follows:
--model Choose the semantic segmentation methods.
--base_model Choose the backbone model.
--dataset The path of the dataset.
--loss The loss function for traing.
--num_classes The number of classes to be segmented.
--random_crop Whether to randomly crop the image.
--crop_height The height to crop the image.
--crop_width The width to crop the image.
--batch_size The training batch size.
--valid_batch_size The validation batch size.
--num_epochs The number of epochs to train for.
--h_flip Whether to randomly flip the image horizontally.
--v_flip Whether to randomly flip the image vertically.
--brightness Randomly change the brightness (list).
--rotation The angle to randomly rotate the image.
--zoom_range The times for zooming the image.
--channel_shift The channel shift range.
--data_aug_rate The rate of data augmentation.
--checkpoint_freq How often to save a checkpoint.
--validation_freq How often to perform validation.
--num_valid_images The number of images used for validation.
--data_shuffle Whether to shuffle the data.
--random_seed The random shuffle seed.
--weights The path of weights to be loaded.
--steps_per_epoch The training steps of each epoch.
--lr_scheduler The strategy to schedule learning rate.
--lr_warmup Whether to use lr warm up.
--learning_rate The initial learning rate.
--optimizer The optimizer for training.
b. start training (All parameters except the selected model are set to default values. Please adjust them as required)
cd 03.Area_segmentation
python train.py --model [segmentation head network] --base_model [backbone]
c. validation
python test.py --model [segmentation head network] --base_model [backbone] --weights [path of *.h5] --dataset [path of validset]
The visualization for these two tasks is the heatmap.
cd 01.Phase_identification
python heatmap --input [folder of data] --model [path of *.pth]
#----------- or -----------
cd 02.Disease_diagnosis
python heatmap --input [folder of data] --model [path of *.pth]
cd 03.Area_segmentation
python visiualization.py --model [segmentation head network] --base_model [backbone] --weights [path of *.h5] --dataset [path of validset]
