Weakly supervised learning on medical images with inaccurate labels

This repository is a PyTorch implementation of influence function applied to InceptionV3 models. The influence function plug-in can also be applied to other deep convolutional neural networks(DCNN) for mislabeled imaging data identification.

Setup

First, install required Python packages with the require.txt in Anaconda Environment.

conda create --name myenv
pip install -r requirements.txt

Next, put the dataset into the data/ folder under weak_supervision directory. Please change BASE_DIR in each py file to your working directory.

Uage

When the ground truth labels exists, we manually create mislabels by flipping labels in a small portion of the training data. Here the percentage of mislabels is 20%, i.e. 0.2 in the commands below. Note that 0.2 can be changed to any number between 0.0 and 1.0.

cd weak_superviion/training
python flipped_label.py 0.2

A DCNN model can be trained with the following command:

cd weak_supervision/
python -m training.inceptionV3 0.2

To calculate the influence of training data on the model:

cd weak_superviion/
python -m influence_function.inceptionV3_influence 0.2

To calculate the percentage of mislabels identified by our method:

cd weak_superviion/
python -m influence_function.inceptionV3_influence 0.2
python mislabel_identification.py

To compare with a previous method, i.e., ensemble learning with multiple majority filter(mmf):

cd weak_superviion/evaluation/mmf/
python mmf_ddsm.py

Reference

Inaccurate labels in weakly-supervised deep learning: Automatic identification and correction and their impact on classification performance.