Building a Mask R-CNN from scratch using TensorFlow and Keras

Tested in TensorFlow 2.4.1
Article about the content of this repository: https://medium.com/@rajkifranciska/building-a-mask-r-cnn-from-scratch-in-tensorflow-and-keras-c49c72acc272

1. Files:

• GenerateToyDataset_fromscratch.ipynb: script to generate toydataset using the masks, and LIDC-IDRI dataset.

• ./masks/: folder containing two mask files to generate toydataset (You need to download the LIDC-IDRI dataset, and have a config file as defined here:https://pylidc.github.io/install.html)

• MaskrCNN.ipynb: creating and training a Mask R-CNN from scratch, using the toydataset. All networks and trainsteps can be observed here.

• MaskrCNN_call.ipynb: Generating and training a new Mask R-CNN, or finetuning saved models can be done here. No functions defined here.

• model_utils.py: training/model functions like: creating network, batchgenerator, trainstep, trainloop, predict, losses are found here.

• utils.py: calculating helper functions + visualization functions are here

If you use my toydataset generation, you will need a path to:

1. the CT slices
2. the normalized CT slices
3. the toydataset, which has to include an images and a masks folder

2. Necessary packages you might not have:

• to work with the nrrd files install pynrrd
• to work with LIDC-IDRI install pylidc

3. Citation

@misc{franciskarajki_maskrcnn_2021,
title={Building a Mask R-CNN from scratch using TensorFlow and Keras},
author={Franciska Rajki},
year={2021},
publisher={Github},
journal={GitHub repository},
howpublished={\url{https://github.com/rajkifranciska/maskrcnn-from-scratch}},
}

Franciska Rajki, Ulyssys Ltd. https://ulyssys.hu/