deep_segment

Segmentation of skin cancer images for submission to ISIC 2018. Base built on RECOD Titans' implementation.

Code is based on https://github.com/learningtitans/isbi2017-part1

Documentation and paper

Arxiv link: https://arxiv.org/abs/1807.04893v1

Preprint of longer version - journal_paper_preprint.pdf available on the repo

Code

IMPORTANT: Please note all default paths in segment.py, pickle_results.py, and ensemble.py

Please create folders 'pickled_results' and 'predicted_masks' within this directory. Change the relevant paths inside pickle_result.py and ensemble.py otherwise.

A folder by name 'saved_models' must exist outside this directory.

Data must be stored as 128x218 images in the base training folder specified inside segment.py

MAIN file is segment.py

After training with segment.py, saved weights are stored in the specified path (please go through segment.py code) Default is '../saved_models/2018_final'

After training a model, you can either evaluate it using 'do_predict=True' in segment.py or use 'pickle_results.py' to do it separately

The pickled results are stored in the folder "pickle_results"

After storing the pickled results, use ensemble.py to combine them all and evaluate them in their original sizes (if masks are available) or/and write the predicted, resized masks out to the folder "predicted_masks"

ISIC_dataset.py, models.py, and metrics.py are used by the main file

post_process.py contains the function for post-processing, but is not used by the main file since the function is also written inside segment.py

Change the numpy random seed to create a different training-validation split and different random weights

Author

Joshua Ebenezer

Date created

July 9th, 2018