Predicting semantic segmentation quality in laryngeal endoscopy images

In this repository, you will find the code associated with predicting the semantic segmentation quality in laryngeal endoscopy images.

Requirements

We used Python 3.10 for our experiments. We heavily rely on the following libraries:

• numpy
• scipy
• TensorFlow (2.16.0) with Keras (3.1.1), for the best experience enable GPU support.

Inside this repository

Creating the IRR dataset

We provide a Jupyter notebook (Create IRR dataset.ipynb) that takes the pre-processed BAGLS dataset (you can download this here), creates a random, but seeded subset from the full dataset (here: 100 samples) and creates three shuffles for assessing inter- and intra-rater reliability.

IoU training data generation

For creating artificial artifacts on the BAGLS training data, we use a multitude of different artifacts focusing on border-pixels, low scale noise and large scale patches using Perlin noise. You find the code in the Jupyter notebook IoU training data generation.ipynb.

Training the classifier

We provide scripts for loading the training data for the IoU prediction deep neural network (DataGenerator.py) and training the actual model (train.py).

How to cite

To be announced