Evaluating Explainable AI Methods to Support Dementia Detection in MRI Scans
Poster No:
1109
Submission Type:
Abstract Submission
Authors:
Devesh Singh1, Stefan Teipel1,2, Martin Dyrba1
Institutions:
1German Center for Neurodegenerative Diseases (DZNE), Rostock, Germany, 2Department of Psychosomatic Medicine, Rostock University Medical Center, Rostock, Germany
First Author:
Co-Author(s):
Stefan Teipel
German Center for Neurodegenerative Diseases (DZNE)|Department of Psychosomatic Medicine, Rostock University Medical Center
Rostock, Germany|Rostock, Germany
German Center for Neurodegenerative Diseases (DZNE)|Department of Psychosomatic Medicine, Rostock University Medical Center
Rostock, Germany|Rostock, Germany
Introduction:
Explainable Artificial Intelligence (XAI) methods have seen increasing adoption in medical applications, particularly for convolutional neural networks to detect e. g. dementia based on MRI. These methods enhance transparency by providing insights into how AI models arrive at their decisions, fostering trust among clinical experts. A recent study highlights the explanation space as a valuable information source, as it yielded a higher group separation than compared to the original input space (Schulz, 2020). However, as relevance or attribution-based XAI methods become more accepted, the need for robust evaluation techniques becomes more urgent (Wang, 2023; Leonardsen, 2024). Different XAI methods can be nonconvergent in their assessment (Fig.1), and without an user-centered evaluation, there is a risk of misleading interpretations. We developed i) three methods for generating explanations, and ii) a qualitative evaluation framework to assess the utility of these XAI explanations in clinical practice.
·Fig.1 Regional volume vs. CNN relevance scores
Methods:
In our study, we used (N≈4000) subjects from six data cohorts, with disease diagnosis of Alzheimer's disease (AD), mild cognitive impairment (MCI) due to AD, (healthy) cognitively normal (CN), and phenotypes of frontotemporal dementia (FTD) subjects. We developed an XAI framework that considers morphological brain signals, i.e., cortical thickness and gray matter volume obtained from FastSurfer, as proxy ground truth signals for disease pathology (Fig.2). Using a clustering method, we considered associations between the morphological signals and the relevance signals generated from a DL model for disease classification. First, in this 'context-enriched explanation space', a neighborhood-based analysis of subjects with longitudinal data available, helped us to create an explanation-by-example method, in particular including possible future cognition trajectories for the query subject. Second, we performed hierarchical clustering for disease subtyping, as a method for model and data explanation by simplification. Third, we developed a custom neuroanatomical ontology and a rule-based methodology to generate textual explanations. Subsequently, through one-to-one qualitative interviews, we evaluated the usefulness of these three proposed methods of model and data explanations with N=5 clinical experts, specifically, neuroradiologists and neurologists.
·Fig.2 Study design for creating explanations for CNNs
Results:
The clustering-based (explanations by simplification) method successfully identified two distinct sub-groups among the subjects: stable subjects and declining subjects. The Kaplan-Meier Survival Analysis revealed a sharp decline in the survival probability for the at-risk subgroup, particularly within the first 12 months of follow-up. Furthermore, mixed-effects modeling demonstrated that declining subjects exhibit a significantly faster rate of cognitive decline (seen from CDR and MMSE scores) compared to stable subjects. Interviews with clinician experts underscored the variability in their explanatory needs. Specifically, radiologists expressed a preference for explanations that focus on pathological descriptions to aid in interpreting imaging results, whereas neurologists favored explanations that assist in diagnostic and therapeutic decision-making.
Conclusions:
Our study demonstrates the effectiveness of a systematic XAI framework in identifying distinct subgroups of subjects, namely stable and declining subjects, using morphological brain signals and convolutional neural network relevance signals. Furthermore, insights from interviews with clinicians emphasize the need for context-specific explanations, with radiologists preferring pathology descriptions and neurologists seeking diagnostic assistance. These findings underscore the importance of robust evaluation techniques and tailored XAI explanations to enhance clinical utility and foster trust in medical AI applications.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Keywords:
Data analysis
Degenerative Disease
Machine Learning
Modeling
Segmentation
STRUCTURAL MRI
Workflows
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
Schulz, M. A., Chapman-Rounds, M., Verma, M., Bzdok, D., & Georgatzis, K. (2020). Inferring disease subtypes from clusters in explanation space. Scientific Reports, 10(1), 12900.
Wang, D., Honnorat, N., Fox, P. T., Ritter, K., Eickhoff, S. B., Seshadri, S., ... & Alzheimer’s Disease Neuroimaging Initiative. (2023). Deep neural network heatmaps capture Alzheimer’s disease patterns reported in a large meta-analysis of neuroimaging studies. NeuroImage, 269, 119929.