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
- 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.
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
| 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 |  |
 |
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