4,068 posters, 219 videos, 101 topics, 69 sessions, 14,365 authors, 6,128 institutions

Alzheimer's Association International Conference 2026

July 12-15, 2026 | London, United Kingdom, and Online

871

Explainable Artificial Intelligence For Automated Amyloid-PET Positivity Classification

View Poster Contact Presenter

Poster Presenter

Sophie Martin

Authors

Sophie A. Martin, Lyduine Collij, Leonard Pieperhoff, Ariane Bollack, Alle Meije Wink, David Vállez-García, Frederik Barkhof, James H. Cole

Affiliations

UCL, UK, Lund University, Sweden, Amsterdam UMC, Netherlands, GE Healthcare; UCL, UK, Barcelona-β Research Centre, Spain; CIBER-BBN, Spain, UCL, UK; Amsterdam UMC, Netherlands

Part of Topic

Motivation:
•Positron emission tomography (PET) imaging can be used to detect amyloid plaques and support early diagnosis of Alzheimer’s disease (AD).
•Artificial intelligence (AI)-based PET classification or quantification [1],[2] can address inter-rater subjectivity and facilitate rapid screening.
•Existing AI models lack interpretability.
•Explainable artificial intelligence (XAI) methods can help to interpret model outputs but produce variable results [3].
•Validating methods is challenging without a 'ground truth’ reference and is often limited group-wise, structural imaging analyses [4].
•There remains a need for assessing the relevance of salient regions on a subject-level.

Aims:

We trained a deep learning model to predict amyloid positivity and validated XAI-generated saliency maps by:

1.Assessing their agreement with subject-level regional distributions of amyloid deposition (alignment)

2.Assessing the effect of removing salient regions on model probability (fidelity)

3.Exploring whether subject-level heatmaps can be used to explain unequivocal cases.

When predicting amyloid positivity, what information does a deep learning model use and which XAI method best captures its reasoning?

Methods:

•1316 baseline static amyloid PET scans from the AMYPAD Prognostic and Natural History Study (PNHS).

•Visual reads were performed by a panel of radiologists [5].

•All participants had a Clinical Dementia Rating score <= 1.

•A 3D convolutional neural network (CNN) was trained to classify between visually amyloid positive and negative individuals.

•A random sample of 80% of the scans were used for model training and the remaining were used as a test set.

•The trained model achieved a test accuracy of 93.2% (89.1% balanced), with 87.5% precision and 81.7% recall.

•Per-tracer test accuracies were 96.7 and 98.4 for FBB and FMT respectively.

Results:

•Model explanations were generated using 10 XAI methods.

•Pearson’s r was used to quantify the correlation between mean saliency and quantification measures in 13 pre-defined ROIs across all test subjects.

•For fidelity, mean saliency was correlated with the change in model probability when that region was occluded from the input PET volume.

Mean LIME and GradCAM saliences produced the highest correlations with SUVr, CL and change in model probability across the 13 ROIs.

•Subject-level explanations reveal distinct patterns of the regions that were most salient to the model’s prediction.

References:

1.Fan et al., Radiology: AI, 2024.

2.Kang et al., Journal of Nuclear Medicine, 2023.

3.Martin et al., Human Brain Mapping, 2026.

4.Wang et al., NeuroImage, 2023.

5.Collij et al., Frontiers in Neurology, 2023.

6.Wolz et al., NeuroImage, 2010.

Comments

Email is required but is not shown publicly.

Sort comments

Loading comments...