Alzheimer's disease detection using explainable AI on PET images. (1st February 2022)

• Record Type: Journal Article
• Authors: Weatheritt, Jack; Palombit, Alessandro; Manber, Richard; Wolz, Robin
• Is Part Of: Alzheimer's & dementia. Volume 17(2021)Supplement 4
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-02-01
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.053831
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Automated detection of Alzheimer's disease features from imaging is a widely worked on machine learning problem, facilitated by the amount of data available through programs such as ADNI. Consequently, "black‐box" neural networks are starting to become viable and accurate approaches to patient classification. However, such methods are typically opaque, and their results are difficult to interpret – limiting their deployment in clinical trials. Our aim through this work is to apply recent advances in explainable deep neural networks and show that our model is indeed sensitive to the underlying metabolic processes. This makes the deep learning prediction process more transparent. Method: We use 520 AC‐PC aligned [18F]‐FDG‐PET scans from the ADNI cohort from AD and healthy control (CG) groups. These are then fed to a deep neural network (ResNet, 3.4 million parameters), which learns to classify each scan. We then validate predictive capabilities of the model on 120 previously unseen scans. Using the validation data, we also calculate a saliency map for each scan. The map represents regions that were important for the model's prediction. For each layer of the neural network, we calculate a cross‐entropy like quantity amongst the activations for a given layer. A group‐wise value is then calculated by averaging the subject maps within each group. The resulting difference highlights which voxels are activated during AD/CG classification. Principal component analysis reveals the regions of spatial location highest variance in the activations. Result: The algorithm scored a mean accuracy of 89.2%. This is comparable with other approaches in the literature. Average saliency maps and their 2nd principal component can be seen in figures 1, 2 and 3. Conclusion: The algorithm is shown to accurately distinguish between AD and CG. By using the tools outlined above, we probe how the model arrives at those conclusions. The directions of maximal variance in the saliency maps shows the importance of the precuneus and hippocampi (increased saliency magnitude). This is corroborated by the difference in the mean CG and AD maps. These regions are known areas of FDG hypometabolism – a result previously identified and available in the literature.