Github, pdf, Notebook

How can active learning solve the lack of labeled images for lung disease recognition

Authors

Koen Desplenter, Quinten Vervynck

Abstract

In this paper we'll use pool-based active machine learning to try to reduce the amount of labeled images that are needed for lung disease recognition. We'll test and compare different query strategies, and experiment with enhancing the images.

Setup

Make sure to you unzip all files in /dataset and /enhanced (some are 7z, watch out).

The minimal directory structure needed to run the notebook should look like this:

.
├── dataset
│   ├── COVID
│   │   ├── images
│   │   └── masks
│   ├── Normal
│   │   ├── images
│   │   └── masks
│   ├── Viral Pneumonia
│   │   ├── images
│   │   └── masks
│   ├── x.npy
│   └── y.npy
├── enhanced
│   ├── COVID
│   │   ├── images
│   │   └── masks
│   ├── Normal
│   │   ├── images
│   │   └── masks
│   ├── Viral Pneumonia
│   │   ├── images
│   │   └── masks
│   ├── x.npy
│   └── y.npy
└── report.ipynb

Note that because the x.npy and y.npy files are already here, you do not need to run any of the initial setup functions in the notebook (which take a long time if you do decide to run them).