Using Deep Learning and Resting-State fMRI to Classify Chronic Pain Conditions

Figures

• 

  FIGURE 1. Group-ICA based parcelation from Biobank UK with (A) 21 and (B) 55 components after the authors excluded the components that are not neuronally driven. (C) Parcelation with 39 parcels and based on a multi-subject dictionary learning process (MSDL). (D) Willard functional atlas with 499 regions of interest.

• 

  FIGURE 2. Flowchart describing the entire process of acquisition, preprocessing, processing, learning, and evaluation. (A) Participants were scanned in a resting-state protocol and (B) functional and structural images were collected. (C) Each subject’s images were preprocessed separately applying standard procedures like time-slicing, realigning, coregistration, artifacts detection, and normalization. (D) A connectivity matrix was created for each subject combining a normalized functional image, a set of confounds, a parcelation, and a connectivity measure. (E) The learning processing starts shuffling the list of subject’s connectivity matrix, we preserved the shuffling index to be replicated with the other inputs set combining the different parcelations and connectivity measures. The next step was to select the independent fold used to compare the different fit models. The remainder data was used in the cross-validation process. All models were trained using a k-fold cross-validation scheme with k = 2 to CNN based models and k = 5 for TOPT models. Inside the cross-validation process, from each testing and validation’s dataset, five thousand synthetic connectivity matrices were created.…

• 

  FIGURE 3. Distribution of values of (left) Balanced Accuracy, (center) Log Loss, and (right) AUC for all 2616 experiments using correlation (blue area) and DTW (orange area). All tails were truncated in maximum and minimum values while the dash lines delimit the quartiles. In the case of Balanced Accuracy and AUC, the distributions for Correlation and DTW differ with p-value < 0.05.

• 

  FIGURE 4. Distribution of values of (left) Balanced Accuracy, (center) Log Loss, and (right) AUC grouped by parcelation size: Biobank_UK_25 (blue), Biobank_UK_100 (Orange), MSDL (green), and Willard (red). All tails were truncated in maximum and minimum values while the dash lines delimit the quartiles.

• 

  FIGURE 5. Enhanced box plot of values of Balanced Accuracy (top), Log Loss (middle), and AUC (bottom) grouped by parcelation size (horizontal axis) and classifier architecture: Ann4BrainsBatch (blue), Ann4Brains (orange), Meszlenyi2017 (green), and TPOT (red).

• 

  FIGURE 6. Local interpretable model-agnostic explanations (LIME) for the best TOPT model listing the 20 features that most contribute to the prediction of (a) a correctly predicted HC, (b) a correctly predicted CP, and (c) a wrongly predicted HC. The bars represent how much each feature (axis Y) added to the prediction in terms of probability. Green bars indicate that the feature adds positively, in direction to the target class, while the red bars indicate that the features are in resistance to the target class.