Attention-Guided Deep Graph Neural Network for Longitudinal Alzheimer’s Disease Analysis - PowerPoint Presentation

1. Attention-Guided Deep Graph Neural Network for Longitudinal Alzheimer’s Disease AnalysisJunbo Ma, Xiaofeng Zhu, Defu Yang, Jiazhou Chen, Guorong WuGraph Deep Learning for Medical ApplicationsSergei VoloboevTUM

2. Motivation

3. Motivation of the AD predictionAlzheimer’s disease (AD) is an irreversible, progressive neurodegenerative diseaseAs pre-symptom interventions may ultimately constitute the best long-term therapeutic strategy, monitoring AD progression is essential for adjusting the therapy plan of the patients to moderate the cognitive declineJanuary 1, 2022Computer Aided Medical ProceduresSlide 3

4. Motivation of the proposed methodBrain network is an efficient way to reveal the complex organizational structures of the brain => usage of graphs is reasonable for the taskSpectral methods for feature extraction are not feasible, because the graph structure has to be fixed across the population. Spatial methods often rely on the random walk process to sample the graph structure.But random walk methods usually lacking the mechanism to optimize the objective of the target task.Authors introduce Attention Guided Random Walk as the way to train random walk for a specific task, and test it for the AD prediction.January 1, 2022Computer Aided Medical ProceduresSlide 4

5. Methodology

6. MethodologyData:Sequence of MRI scans with the final label for each patient Each scan initially represented as a fully-connected graph of 148 cortical regions, and then the edges are filtered with disparity filterJanuary 1, 2022Computer Aided Medical ProceduresSlide 6

7. MethodologyTask:Using a temporal sequence of brain graphs, assign label to each one (AD/non-AD).Due to the lack of data, the authors reduce this task to classifying the entire sequence.(examples are created by me, because the dataset is private)January 1, 2022Computer Aided Medical ProceduresSlide 7

8. Vanilla Random WalkJanuary 1, 2022Computer Aided Medical ProceduresSlide 8Just assign equal weights to neighbors and choose random for the next step. Walk until the maximum length is achieved.How to adapt it to a specific task and summarize the route information?

9. Attention-Guided Random Walk Module (the proposed way to create an embedding of a single graph)Step 0The algorithm starts by taking a random node as a starting point and initializing the attention weights A0 to the uniform distribution across all the nodes.January 1, 2022Computer Aided Medical ProceduresSlide 9

10. Attention-Guided Random Walk Module Step 1The current rank vector   rt is then calculated by   rt=at⊙wt, where   ⊙  represents the element-wise production and   wt  is the edge weights vector.Then, the rank vector is normalized as a probability distribution, where the sum of ranks is 1. Based on this rank vector, the next node to walk on is selected randomly.January 1, 2022Computer Aided Medical ProceduresSlide 10

11. Attention-Guided Random Walk Module Step 2 and 3On each time step, the current attention at and features of the next node are projected to a hidden space by linear layers with parameters θ1  and   θ2. These projections are concatenated and projected again with weights θ3. The result of these operations is called step state sstThen, the step state is fed to the first LSTM (step LSTM)  as an input.January 1, 2022Computer Aided Medical ProceduresSlide 11

12. Attention-Guided Random Walk Module Step 4New attention weights are calculated as at+1 = fa(h_{t})The produced hidden state of LSTM on each step is then used to predict the label of the graph. January 1, 2022Computer Aided Medical ProceduresSlide 12

13. Attention-Guided Random Walk Module Step 5Back to step 1, while maximum route length is not achieved.January 1, 2022Computer Aided Medical ProceduresSlide 13

14. Attention-Guided Random Walk Module Step 6At the last step T: the reward equals to 1 if the step LSTM correctly predicts the graph label and 0 otherwise. The reward is then used in loss calculation. Hidden state T used as graph embedding January 1, 2022Computer Aided Medical ProceduresSlide 14

15. Attention-Guided Random Walk Module loss(μ∈[0,1] is slowly growing to 1 according to the number of steps)January 1, 2022Computer Aided Medical ProceduresSlide 15

16. Sequence Processing ModuleThe sequence of graph embeddings from AGRW is further processed by sequence LSTM (another one) with global attention to predict the final label. The prediction loss of the sequence processing module is combined with the AGRW loss:January 1, 2022Computer Aided Medical ProceduresSlide 16

17. Proposed model frameworkJanuary 1, 2022Computer Aided Medical ProceduresSlide 17

18. Experimental setup

19. Data preparation (Graphs generation) The proposed model is tested on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset [1]. MRI images are preprocessed via this pipeline:The image is divided into 148 cortical regions;Fiber tractography is applied to construct the 148 × 148 connection matrices, which represents the brain networks;Then, disparity filtering is applied to make the brain networks sparse [2];There are 325 brain networks in total from 94 subjects (70/24 train-test split). The sequence lengths of the subjects vary from 1 to 6, with an average of 3.45.[1] Jack Jr, C. R., Bernstein, M. A., Fox, N. C., Thompson, P., Alexander, G., Harvey, D.: The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. J. Magnet. Reson. 27(4), 685–691 (2008). Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine[2] Serrano, M.A., Boguna, M., Vespignani, A.: Extracting the multiscale backbone of complex weighted networks. Proc. Natl. Acad. Sci. 106(16), 6483–6488 (2009)January 1, 2022Computer Aided Medical ProceduresSlide 19

20. Data preparation (Node features)There are 3 attributes for each brain region:Cortical thickness derived from T1-weighted MRI images,FDG level derived from 18F-FDG PET images, Amyloid β level.All the attributes are rescaled to [0,1]January 1, 2022Computer Aided Medical ProceduresSlide 20

21. BaselinesSVM over concatenated nodes features of the first scan is used as a baseline model in the paper.  As features for each node, they use:1. The node degree2. Local efficiency3. Cluster coefficient4. 3 node attributes from the dataJanuary 1, 2022Computer Aided Medical ProceduresSlide 21Attention-Guided Random Walk Module over first scans used as another baseline to show the efficiency of the proposed way of feature extraction

22. HyperparametersOnly one combination of hyperparameters for the model was tested:Random walk length: 150Step-LSTM hidden size: 128Sequence-LSTM hidden size: 256The dimensions of the weights θ1 ,θ2 and θ3 for step state are not mentioned.The network is trained by the Adam optimizer until 500 iterations, or no loss decline over 50 iterations.January 1, 2022Computer Aided Medical ProceduresSlide 22

23. Results

24. ResultsTo show the efficiency of a single AGRW module compared to the SVM baseline, the authors use AGRW embedding for the first scan for each subject to predict the label. January 1, 2022Computer Aided Medical ProceduresSlide 24

25. InterpretationThe visualization of the attention vector from the last step. January 1, 2022Computer Aided Medical ProceduresSlide 25

26. InterpretationThe node frequency of the top 10 nodes in each of the last 30 steps.January 1, 2022Computer Aided Medical ProceduresSlide 26

27. State of the art[1] Cui, R., Liu, M., & Alzheimer's Disease Neuroimaging Initiative. (2019). RNN-based longitudinal analysis for diagnosis of Alzheimer’s disease. Computerized Medical Imaging and Graphics, 73, 1-10.[2] Pölsterl, S., Aigner, C., & Wachinger, C. (2021, September). Scalable, Axiomatic Explanations of Deep Alzheimer’s Diagnosis from Heterogeneous Data. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 434-444). Springer, Cham.January 1, 2022Computer Aided Medical ProceduresSlide 27RNN-based longitudinal analysis for diagnosis of Alzheimer’s disease [1](2019, CNN for feature extraction, BGRU for temporal component)this paper(2020)Scalable, Axiomatic Explanations of Deep Alzheimer’s Diagnosis from Heterogeneous Data [2](2021, Wide and Deep PointNet) accuracy, %91.3393.67balanced accuracy of 94.2

28. My review (strengths)In my opinion, the paper tackles its difficult yet important task with decent results in terms of accuracy and interpretability. Using attention both in feature extraction and temporal components gives an easy way to extract important regions for the task.Proposed Attention-Guided Random Walk Module gives an ability to process big graphs as it uses only local information. January 1, 2022Computer Aided Medical ProceduresSlide 28

29. My review (weaknesses)However, there are multiple research-conducting issues. I identified the following problems in the paper:January 1, 2022Computer Aided Medical ProceduresSlide 29Conceptual:The overcomplicated approach without an ablation study

30. My review (weaknesses)However, there are multiple research-conducting issues. I identified the following problems in the paper:January 1, 2022Computer Aided Medical ProceduresSlide 30Data-related:Is the model capable of predicting future disease, or can it only distinguish between damaged and non-damaged brains?Possible overkill for 94 sequencies classificationConceptual:The overcomplicated approach without an ablation study

31. My review (weaknesses)However, there are multiple research-conducting issues. I identified the following problems in the paper:January 1, 2022Computer Aided Medical ProceduresSlide 31Data-related:Is the model capable of predicting future disease, or can it only distinguish between damaged and non-damaged brains?Possible overkill for 94 sequencies classificationExperimental:No comparison with previous methods for the taskNo comparison between multiple models for sequence processing and graph features extraction (and with vanilla random walk too)The code is not sharedNo motivation and/or experimental data for chosen hyperparametersPredicting only 1 last label for the sequence of scans may result in no sense of temporal component of the model, but the authors avoided this conclusion when they trained baselines on the first scan, not the last, which seems to be more relevant to predicting the final labelConceptual:The overcomplicated approach without an ablation study

32. Take home messageRandom walk approach might be useful for big graphs It is also an interpretable way for creating graph embeddingsLack of comparison with SOTAs and the ablation study makes any approach questionable, even with a good result for the taskTesting a model with many parameters on a small dataset without pretraining is questionable, tooUsing multiple datasets for testing a model is also a good way to make a research more trustworthyJanuary 1, 2022Computer Aided Medical ProceduresSlide 32