A knee cannot have lung disease: out-of-distribution detection with in-distribution voting using the medical example of chest X-ray classification

This repository includes the official implementation of "A knee cannot have lung disease: out-of-distribution detection with in-distribution voting using the medical example of chest X-ray classification" by Alessandro Wollek, Theresa Willem, Michael Ingrisch, Bastian Sabel and Tobias Lasser. This method is an effective OOD detector for multi-label classification problems.

Pre-trained Models

The pre-trained models can be found in models/.

Test Results

All test results are calculated in the according Jupyter notebooks in notebooks/.

Training

To train the models from scratch you need the following data sets and modify the respective data loaders according to the file structure.:

• Chest X-Ray 14
• IRMA
• MURA
• Bone Age
• ImageNet