Skip to content

Public code repository for survey deployment (React.js/Node.js/MongoDB) and data processing (Python/Jupyter) for the journal paper "The Explainability Paradox: Challenges for xAI in Digital Pathology" (2022) FGCS Special Issue on xAI in healthcare

License

Notifications You must be signed in to change notification settings

theodore-evans/xai-in-digital-pathology

Repository files navigation

The explainability paradox: Challenges for xAI in digital pathology CODE REPOSITORY

DOI

Online questionnaire data and data processing, supporting:

Evans, T., Retzlaff, C., Geißler, C., Kargl, M., Plass, M., & Müller, H. et al. (2022). The explainability paradox: Challenges for xAI in digital pathology. Future Generation Computer Systems. doi: 10.1016/j.future.2022.03.009

Data analysis is available in the accompanying Jupyter notebook

Questionnaire contents

  • User profiling questions, collecting data on usage of and familiarity with AI applications in pathology, and with machine learning in general
  • 7 example implementations of explainability methods on a sample Ki-67 app output, with 4 Likert-scale feedback questions to evaluate intelligibility, informativeness and value to user.

Viewing the survey

git clone https://github.com/theodore-evans/xai-in-digital-pathology.git
cd xai-in-digital-pathology/frontend
npm i
npm start

Open http://localhost:3000/ in your web browser

Explanation examples

Name Description
Saliency Map (Local) Show the most relevant pixels for the classification of a selected annotation
Saliency Map (Global) Show the most relevant pixels for the positive classifications within this region of interest
Concept Attribution Show the most important features attributed to positive classifications
Prototypes Show prototypical positively and negatively classified annotations within this region
Counteractuals (One-axis) Show generated examples interpolating between positive and negative examples, showing model classifications for each
Counteractuals (Two-axis) Show generated examples changing in two principal factors of variation, showing model classifications for each
Trust Scores Display low-confidence annotations for review

Additional info

Sample Ki-67 model: PathnoNet, trained for 20 epochs on the training set of SHIDC-B-Ki-67-V1.0 and demonstrated with the test set of the same dataset.

GradCAM heatmap generated using Neuroscope-1.0

Example interpolations mocked up using DiffMorph

All other graphics created with GIMP

For more detail on example creation, please refer to Method > Questionnaire design in Evans et Al (2022)

This survey is build in React.js using survey.js. The project was adapted from SurveyJS for React quickstart project Public code repo for FGCS Special Issue on xAI in healthcare

About

Public code repository for survey deployment (React.js/Node.js/MongoDB) and data processing (Python/Jupyter) for the journal paper "The Explainability Paradox: Challenges for xAI in Digital Pathology" (2022) FGCS Special Issue on xAI in healthcare

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •