This repo contains implementation code for the deep multi-instance learning (MIL) method proposed within the dissertation entitled: Deep Multi-Instance Learning for Automated Pathology Screening in Frontal Chest Radiographs. The method is implemented within Python using Tensorflow to express model architectures and faccilitate network training.
NOTE: This implementation code has been tested on a machine runnning python 2.7 and Tensorflow 1.4. However, it should run seamlessly on python 3.x distros and later versions of Tensorflow.
- Download an image dataset for which binary classification is suitable. For example, the Kaggle Pneumonia X-ray Collection
- Partition the dataset into a positive and a negative class. Place the positive items into
data/datasets/raw/d_1
and the negative items intodata/dataset/raw/d_0
- Prepare the data and launch MIL training by running:
source launch_MIL.sh
from the terminal. - At the conclusion of training, the resulting MIL model, as well as its performance on a test set of the data, is stored
data/checkpoint/stage_2
Code within this repository provided under the GPLv3 License.