Multi-modal Generation of Counterfactual Baselines for Contrastive Explanations

• Applicant: Professor Dr. Christian Wachinger
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 565014922

Deep neural networks hold transformative potential in various domains, especially in medicine. Yet, clinical acceptance and current regulations require transparent decision-making, not natively provided by opaque neural nets. Explainable AI (XAI) aims to increase their transparency. Contrastive explanations clarify why a model made a particular decision by comparing the actual to a counterfactual outcome. They are, therefore, considered desirable, as they mirror human reasoning and reflect the medical diagnostic process by contrasting possible diagnoses. However, counterfactual outcomes cannot be observed and need to be synthetically generated. The counterfactual can then be set as the baseline in XAI methods, representing missingness, to obtain contrastive explanations. We will focus on Shapley values due to their attractive theoretical properties. This project aims to advance XAI in the relevant medical application of dementia prediction, with patients expected to triple by 2050 due to an aging society. We will consider the differential diagnosis of dementia, a classification task, and the prediction of the time to dementia onset, modeled with survival analysis. The prediction network must handle multi-modal data, including neuroimaging (MRI and PET scans) and clinical information. Uniquely, we will reconstruct cortical surfaces and map relevant information like cortical thickness and metabolic activity onto the surfaces. The neural networks will then take surfaces, represented as triangular meshes, instead of images as input. Learning on surfaces will distinguish this project from existing work and require novel developments based on geometric deep learning. The benefit will be that visualizing explanations on surfaces will be considerably clearer than on 3D volumes. Other technical challenges in this project include the need to approximate Shapley values, which are computationally infeasible for high-dimensional data. We will use an approximation that requires the design of a probabilistic prediction network. Further, we will use denoising diffusion models for the baseline generation, which typically expect paired training data. Since we will have unpaired training data, we will use the trained prediction models as discriminators and formulate a cycle consistency loss. This will ensure that the generation fulfills the target condition and preserves patient-specific information. To achieve these goals, we will combine many dementia datasets and use a distinctive in-house dataset to have sufficient data for training and evaluation. The explanations will be rigorously evaluated using synthetic and medical data. With our clinical collaborators, we will evaluate their impact in a clinical user study. This project will potentially set new standards in the application of AI in medical diagnostics, offering clearer, more accessible insights into its decision processes, thus bridging the gap between complex AI models and clinical usability.

• DFG Programme: Research Grants