Special Focus Session
SF 15b - Validity of explainable AI and model trustworthiness
Lectures
1
Chairperson's introduction
05:00Federica Zanca, Leuven / BE
2
Assessing the trustworthiness of saliency maps as a tool to increase clinical explainability of AI radiology models
25:00Kostas Marias, Heraklion, Crete / GR
3
Generation of visual counterfactual explanation for classification problems in (medical) imaging
25:00Gaspard D'Assignies, Nantes / FR
4
Quantifying explainability: does it improve clinical acceptance of AI?
25:00Alberto Traverso, Milan / IT
5
Panel discussion: Who should care about explainability? Is explainability the best choice, or should we invest resources in developing inherently interpretable and transparent models?
10:00Panel discussion: Who should care about explainability? Is explainability the best choice, or should we invest resources in developing inherently interpretable and transparent models?
5 min
Chairperson's introduction
Federica Zanca, Leuven / Belgium
25 min
Assessing the trustworthiness of saliency maps as a tool to increase clinical explainability of AI radiology models
Kostas Marias, Heraklion, Crete / Greece
1. To present the most popular saliency map techniques.
2. To demonstrate current applications in radiology (e.g. highlighting regions of attention indicating pathology).
3. To describe how they can be used within the radiology workflow along with their current limitations.
2. To demonstrate current applications in radiology (e.g. highlighting regions of attention indicating pathology).
3. To describe how they can be used within the radiology workflow along with their current limitations.
25 min
Generation of visual counterfactual explanation for classification problems in (medical) imaging
Gaspard D'Assignies, Nantes / France
1. To understand this type of explanation and the link with the radiologist's practice.
2. To learn about methods with slightly different outputs that bring complementary information.
3. To be aware of the risks of each of these methods.
2. To learn about methods with slightly different outputs that bring complementary information.
3. To be aware of the risks of each of these methods.
25 min
Quantifying explainability: does it improve clinical acceptance of AI?
Alberto Traverso, Milan / Italy
1. To identify the most common explainability algorithms.
2. To list the clinical applications of such algorithms.
3. To understand the impact of such algorithms in clinical decision-making.
2. To list the clinical applications of such algorithms.
3. To understand the impact of such algorithms in clinical decision-making.
10 min
Panel discussion: Who should care about explainability? Is explainability the best choice, or should we invest resources in developing inherently interpretable and transparent models?