Predict cancer segmentations using DNN

Framework

Tensorflow

Models

• U-Net
• MulmoU-Net

Install

pip3 install git+https://github.com/yoshihikoueno/DNNCancerAnnotator.git@master

Usage

Available commands

python3 -m annotator -h

# usage: python3 -m annotator [-h] {train,evaluate,extract_all,generate_tfrecords} ...
# 
# DNNAnnotator: CLI interface
# 
# positional arguments:
#   {train,evaluate,extract_all,generate_tfrecords}
#                         command
#     train               Train a model with specified configs.
#     evaluate            Evaluate a model with specified configs
#     extract_all         extract indivisual images (TRA, ADC, etc...) from the screenshots
#     generate_tfrecords  Generate TFRecords
# 
# optional arguments:
#   -h, --help            show this help message and exit

Train

Command:

python3 -m annotator train

Options:

python3 -m annotator train -h

# usage: python3 -m annotator train [-h] --config CONFIG [CONFIG ...] --save_path SAVE_PATH --data_path DATA_PATH [DATA_PATH ...]
#                                   --max_steps MAX_STEPS [--early_stop_steps EARLY_STOP_STEPS] [--save_freq SAVE_FREQ] [--validate]
#                                   [--val_data_path VAL_DATA_PATH [VAL_DATA_PATH ...]] [--visualize] [--profile]
# 
# Train a model with specified configs.
# 
# This function will first dump the input arguments,
# then train a model, finally dump reults.
# 
# optional arguments:
#   -h, --help            show this help message and exit
#   --config CONFIG [CONFIG ...]
#                         configuration file path
#                             This option accepts arbitrary number of configs.
#                             If a list is specified, the first one is considered
#                             as a "main" config, and the other ones will overwrite the content
#   --save_path SAVE_PATH
#                         where to save weights/configs/results
#   --data_path DATA_PATH [DATA_PATH ...]
#                         path to the data root dir
#   --max_steps MAX_STEPS
#                         max training steps
#   --early_stop_steps EARLY_STOP_STEPS
#                         steps to train without improvements
#                             None(default) disables this feature
#   --save_freq SAVE_FREQ
#                         interval of checkpoints
#                             default: 500 steps
#   --validate            also validate the model on the validation dataset
#   --val_data_path VAL_DATA_PATH [VAL_DATA_PATH ...]
#                         path to the validation dataset
#   --visualize           should visualize results
#   --profile             enable profilling

Evaluate

Command:

python3 -m annotator evaluate

Options:

python3 -m annotator evaluate -h

# usage: python3 -m annotator evaluate [-h] --save_path SAVE_PATH --data_path DATA_PATH [DATA_PATH ...] --tag TAG [--config CONFIG]
#                                      [--avoid_overwrite] [--export_path EXPORT_PATH] [--export_images] [--export_csv]
#                                      [--min_interval MIN_INTERVAL]
# 
# Evaluate a model with specified configs
# 
# for every checkpoints available.
# 
# optional arguments:
#   -h, --help            show this help message and exit
#   --save_path SAVE_PATH
#                         where to find weights/configs/results
#   --data_path DATA_PATH [DATA_PATH ...]
#                         path to the data root dir
#   --tag TAG             save tag
#   --config CONFIG       configuration file path
#                             None (default): load config from save_path
#   --avoid_overwrite     should `save_path` altered when a directory already
#                             exists at the original `save_path` to avoid overwriting.
#   --export_path EXPORT_PATH
#                         path to export results
#   --export_images       export images
#   --export_csv          export results csv
#   --min_interval MIN_INTERVAL
#                         minimum interval in steps between evaluations.
#                             Checkpoints which are less than `min_interval` steps away
#                             from the previous one will be disregarded.