1

Enhancing Functional Connectome Fingerprinting

with Deep Learning

Yashaswini

B.Tech (Electrical Engineering)

and

M.Tech (Signal Processing and Machine Learning)

20EE38020

Supervisor

Dr. Sanjay Ghosh

Assistant Professor,

Department of Electrical Engineering

30 April 2025

​

1

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

2

Brain Fingerprinting: Overview

• We have unique brains with unique
• Structural, functional and dynamical properties
• Shaped by genetics, environment and experiences

​

• Uniqueness is characterized by differences in
• Learning
• Memory
• Cognitive processing

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 2

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

3

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 3

Brain Fingerprinting: Data Sources

Structural Connectivity (MRI) Functional Connectivity (fMRI) Electrophysiological (MEG/EEG)

fMRI: functional Magnetic Resonance Imaging; MEG/EEG: Magneto/ElectroEncephaloGram

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 4

Problem Statement

Develop a robust brain fingerprinting framework to identify individuals

based on their functional connectomes (FC) (derived from fMRI data)

using signal processing and deep learning techniques

Pearson

Correlation

​

and

further processing

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

5

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 5

Literature Review: FC Fingerprint Extraction

Simple Correlation-Based Sparse Dictionary Learning Deep Learning

FC Fingerprint = FC FC Fingerprint = Residual FC FC Fingerprint = Refined FC

​

Reference: https://www.nature.com/articles/nn.4135 Reference:https://pmc.ncbi.nlm.nih.gov/articles/PMC6865523/ Reference:https://arxiv.org/abs/2006.09928

​

​

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 6

Proposed Solution Approach

• Limitations of previous deep learning solutions:
• Applying autoencoder directly on fMRI time series is computationally inefficient (approx. 3.75 M parameters)
• It is less interpretable

​

• Idea: Instead of common neural activities, what if we separate out most common neural connections across subjects

​

• Representing common neural connections
• Baseline: Group-Average Functional Connectomes (no training), Finn (2015)
• Convolutional Autoencoder (approx. 33 K parameters)

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 7

Proposed Solution Approach: Source data to Brain Fingerprint

fMRI Data Extraction and Preprocessing

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 8

Proposed Solution Approach: Source data to Brain Fingerprint

Functional Connectome Generation

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 9

Proposed Solution Approach: Source data to Brain Fingerprint

Training Convolutional Autoencoder to get Reconstructed FC

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 10

Proposed Solution Approach: Source data to Brain Fingerprint

Subtract Reconstructed FC from Original FC to get Residual FC

“Group” FC

“Individualistic” FC

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 11

Proposed Solution Approach: Source data to Brain Fingerprint

Sparse Dictionary Learning Framework to get Refined FC

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 12

Proposed Solution Approach: Why Convolutional Autoencoder (CAE)

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 13

Proposed Solution Approach: Two sample example

ROI

ROI

ROI

ROI

ROI

ROI

ROI

ROI

ROI

ROI

Correlation →

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 14

Proposed Solution Approach: Data Extraction and Preprocessing

fMRI data from subset of HCP (Human Connectome Project) dataset

4D (x, y, z, t) : (91 x 109 x 91, t)

for 339 subjects across resting state and tasks: motor, emotion, working memory, language etc.

Glasser Parcellation:

360 Regions of interest (ROI) over 12 functional networks

• HCP Minimal Preprocessing Pipeline

Blood Oxygen Level Dependent (BOLD) time series signals for each ROI extracted

FC generated using Pearson Correlation between time series

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 15

Glasser Parcellation

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 16

Proposed Solution Approach: CAE Construction

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 17

Proposed Solution Approach: Training CAE

• Trained on Resting-state FC

​

• Loss function = MSE Reconstruction error

​

• Reconstructed output = group-wise FC

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 18

Proposed Solution Approach: SDL Theory (contd.)

Residual FC Group-wise FC using Sparse Coding Refined FC

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 19

Proposed Solution Approach: SDL Theory (contd.)

Residual FC Group-wise FC using Sparse Coding Refined FC

L : model parameter to control the sparsity

level of dictionaries.

D ∈ ℝ^m × K : dictionaries

K : the size of dictionaries

X = [x₁, x₂, …, x_K] ∈ ℝ^K × n : representation matrix

∥ · ∥₀, ∥  · ∥_F : 0 and Frobenius norm

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 20

Proposed Solution Approach: SDL Implementation

• k-SVD algorithm used for implementing Sparse Dictionary Learning

​

Initialize Dictionary

​

​

Sparse Coding using OMP Algorithm

​

Solve for X while keeping D fixed such that X is L-sparse

​

​

​

​

Dictionary update using SVD

​

D is refined to represent residuals keeping the sparsity of X fixed

​

​

Repeat until convergence

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 21

Baseline Solution Approach 1: Finn et. al. (2015)

Used for fingerprinting

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 21

Baseline Solution Approach 2: Group Average Functional Connectomes

Averaged across all

subjects in dataset

Group Average

​

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 23

Individual Identification Accuracy

• Goal: Given a target session (i.e., resting-state), identify the participant from a task session(i.e, Motor, Working Memory, Emotion) based on their refined FC.

​

​

​

​

​

​

​

​

• Accuracy Evaluation: If the matched FCs belong to the same participant, accuracy = 100% else 0%

​

​

​

​

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Final Accuracy of Proposed Method

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 24

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Accuracy with respect to parameters (K and L)

Accuracies for different K and L values: On an average 10% higher on Motor Task

Accuracy →

Accuracy →

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 25

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Accuracy with respect to parameters (K and L)

​

Accuracies for different K and L values: On an average 10% higher on Working Memory Task

Accuracy →

Accuracy →

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 26

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Accuracy with respect to parameters (K and L)

​

Accuracies for different K and L values: On an average 10% higher on Emotion Task

Accuracy →

Accuracy →

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 27

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Ablation Analysis - Motor Task

• Largest Accuracy Decrease:
• Frontoparietal network contribute to motor planning and control
• Posterior-Mu likely reflects sensorimotor integration
• Visual2 may support visuomotor coordination.

​

• Accuracy increase: Excluding Cingulo-Oper and Somatomotor
• These subnetworks may introduce redundancy or noise for this task or are not important for individual characterization

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 28

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Ablation Analysis - Working Memory Task

• Largest Accuracy Decrease:
• Language critical for working memory, especially for tasks requiring symbolic reasoning
• Visual2 important for maintaining spatial or visual details in working memory

​

​

• Minimal Impact on Accuracy:
• Excluding Auditory, Cingulo-Oper, Frontopariet, or Ventral-Mult

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 29

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Ablation Analysis - Emotion Task

• Largest Accuracy Decrease:
• Frontoparietal essential role in emotional regulation and integration
• Visual2 Involved in visuospatial processing of emotional stimuli, particularly facial expressions.

​

• Accuracy Increase:
• Orbito-Affective, Ventral-Mult, Auditory, Visual1, results in slight accuracy increases or negligible.
• These networks dont play a central role in individual fingerprinting for the emotion task

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 30

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Results: Training on other tasks

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 31

1. Rest b) Motor

c) Working Memory d) Emotion

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Future Work

• Other measures of functional connectivity - mutual information, various distance measures like Euclidean distance

​

• Integrate Convolutional Autoencoder and Sparse Dictionary Learning into a single framework
• k-Sparse Autoencoder
• Deep Dictionary Learning

​

• Using the FC Fingerprints obtained for downstream tasks
• Predictive cognitive functions (Ex: fluid intelligence, problem solving skills)
• Individualized diagnosis of neurological disorders

​

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 32

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Conclusion

• Functional Connectivity is a good measure for individualized brain connectivity analysis

​

​

• Functional Connectome Fingerprinting has applications in personalized treatment of neurological disorders and also in cognitive state decoding

​

​

• As the number of fMRI datasets increase, there are chances of missing labels. Functional Connectome Fingerprinting can be a good way to tackle this issue

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 33

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

References

[1] Finn ES, Shen X, Scheinost D, Rosenberg MD, Huang J, Chun MM, Papademetris X, Constable RT. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat Neurosci. 2015 Nov;18(11)

​

[2] Cai B, Zhang G, Hu W, Zhang A, Zille P, Zhang Y, Stephen JM, Wilson TW, Calhoun VD, Wang YP. Refined measure of functional connectomes for improved identifiability and prediction. Hum Brain Mapp. 2019 Nov

​

[3] Cai, Biao, et al. "Functional connectome fingerprinting: identifying individuals and predicting cognitive functions via autoencoder." Human Brain Mapping 42.9 (2021): 2691-2705.

​

Publication

• Yashaswini and S. Ghosh, "Enhancing Functional Connectome Fingerprinting using Deep Learning," IEEE Transactions on Medical Imaging. (in-preparation)

30/04/2025 Enhancing Functional Connectome Fingerprinting using Deep Learning 34

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Thank you!

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Appendix

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

HCP minimal preprocessing pipeline

• Gradient Distortion Correction: Compensates for MRI gradient coil-induced distortions.
• Head Motion Correction: Realigns volumes to correct for subject movement (FIX-based ICA denoising).
• Image Distortion Correction: Addresses echo-planar imaging (EPI) distortions.
• Spatial Normalization: Aligns images to MNI standard space for cross-subject consistency.
• Intensity Normalization: Standardizes signal intensity across scans.
• Temporal Filtering:
• Detrending: Removes linear trends in BOLD signals.
• Band-pass filter (0.01–0.25 Hz): Reduces low-frequency drift and high-frequency noise.
• Spatial Smoothing: Omitted per Finn et al. (2015) findings on minimal impact on fingerprinting accuracy.

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

AE vs Conv-AE Network Parameter Calculations

• AE (architectured as mentioned in the reference paper):
• d-500-500-2000-10-2000-500- 500-d where d = number of time points
• For resting state fMRI, d = 1200
• Number of params = 1200×500+500 + 500×500+500 + 500×2000+2000 + 2000×10+10 + 10×2000+2000 + 2000×500+500 + 500×500+500 + 500×1200+1200 = 3,747,210

​

• Conv AE (proposed architecture):
• Number of params = 1×16×3×3 + 16 + 16×32×3×3 + 32 + 32×64×3×3 + 64 + 64×32×2×2 + 32 + 32×16×2×2 + 16 + 16×1×2×2 + 1 = 33,649

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Orthogonal Matching Pursuit Algorithm

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Dictionary update using SVD

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Transposed Convolution

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Why reconstructed FCs represent group-wise connections?

• Trained across subjects - captures patterns common across the group

​

• CAE compresses and reconstructs inputs - inherently identifies most salient connections

​

• Trained using reconstruction loss - penalizes deviations from population-level patterns i.e, effectively computes a group-average like representation but with more sensitivity to non-linear representations

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

How SDL acts as a feature magnifier ?

• The residual FC contain subject-specific connectivity features - but noisy since high dim

​

• SDL acts a pattern compressor - learns a dictionary of “atoms” that frequently appear across subjects. Each residual FC sparse combination of these atoms

​

• Enforcing sparsity => amplifies discriminative features, suppresses variations

​

Indian Institute of Technology Kharagpur

Department of Electrical Engineering

Low performance of Finn et. al. (2015)

Indian Institute of Technology Kharagpur

Department of Electrical Engineering