Joint Weakly and Semi-Supervised Deep Learning for Localization and Classification of Masses in Breast Ultrasound Images

This is the code for "Joint Weakly and Semi-Supervised Deep Learning for Localization and Classification of Masses in Breast Ultrasound Images".

This is based on a Tensorflow implementation of Faster R-CNN (https://github.com/smallcorgi/Faster-RCNN_TF), which is adopted as the mass detector in the proposed general framework as a choice. Some part of the following descriptions might be a repetition of those in the repository.

Dependency

• Python 2.7.12
• Tensorflow 1.12
• Cython 0.27.3
• easydict 1.7
• pyyaml 3.12
• scikit-image 0.14.2

Installation

• Building Cython codes

cd $ROOT/code/lib
make

Testing a Model

1. Download available trained models. [OneDrive]
2. Run $ROOT/code/main/test.py with appropriate input arguments, including the path for the downloaded model.

Training a Model

1. Download ImageNet pretrained models
   • VGG-16 : https://github.com/smallcorgi/Faster-RCNN_TF
   • ResNet-18,34,50,101 : [OneDrive]
2. Run $ROOT/code/main/train.py (combined mini-batch) or $ROOT/code/main/train_alter.py (alternating mini-batches) with appropriate input arguments, including the path for the downloaded pretrained model.

SNUBH Dataset

We provide sample images corresponding to those in Fig. 6 of our paper. The original result images also can be found in $ROOT/code/qual_res/fig6.

Citation

@article{shin_tmi19,
  author = {S. Y. {Shin} and S. {Lee} and I. D. {Yun} and S. M. {Kim} and K. M. {Lee}},
  journal = {IEEE Transactions on Medical Imaging},
  title = {Joint Weakly and Semi-Supervised Deep Learning for Localization and Classification of Masses in Breast Ultrasound Images},
  year = {2019},
  volume = {38},
  number = {3},
  pages = {762-774},
  doi = {10.1109/TMI.2018.2872031},
  month = {March},
}