Janis Zenkner

Advisors
Robert Richer (M.Sc.), Prof. Dr. Björn Eskofier, Thomas T. Liu (Ph. D.)

Duration
02 / 2022 – 08 / 2022

Abstract

“You are what you think”. While this motto was picked countless times in literature, it must rather be reformulated to make it accurate in a medical way: “You can be identified by how you think”. In fact, Functional Connectome Fingerprinting (FCF) allows to identify single individuals by their brain connectivity, i.e., the connectome, and correlation patterns between several hundreds of regions of interest (ROI) extracted from functional Magnetic Resonance Imaging (fMRI) data [1].

Over the past couple of years, numerous approaches to boost the FCF identification rate have been proposed leading to almost perfect identification accuracies [2][3][4]. Moreover, multiple approaches using the concept of FCF to predict disease progression and behavioral traits have been published [5].

To allow a proper evaluation of the trained machine learning algorithms, the dataset is split into three subsets: training, validation, and test set. The data, generally provided by the Human Connectome Project [6], consists of four fMRI scans per participant that were conducted on two subsequent days. So far, the data were typically split along the scan dimension, i.e., the training set contains the two scans of first day, the validation set the first session of the second day, and the test set the second scan of the second day [2][3][4]. Consequently, each set contains scans of the same subject leading to a data leakage. Yet, this also means that current results must be seen skeptically and might not be applicable to scenarios outside the specific dataset used in the respective publication.

Siamese Neural Networks (SNNs) pose an alternative approach to FCF. SNNs are neural network architectures that contain two or more identical subnetworks [7]. Consequently, SNNs learn to extract similarities in the input data [7]. SNNs also have been used in the FCF setting by predicting whether the input samples are from the same or different subjects [5][8][9]. Yet again, the session split was used. However, SNNs also have the potential to perform well on a data split that prevents the data leakage problem without losing a lot of potential training data, i.e., splitting the dataset along the subject dimension.

The goal of this master’s thesis – that is conducted in cooperation with University of California, San Diego – is to implement SNN architectures that perform well on the subject split. In addition to these architectures, a benchmarking algorithm will be developed to allow further validation. Moreover, the SNN will then be interpreted to try to understand which features the network uses to identify whether two samples are from the same subject or not. Finally, the SNN results are compared to published Neural Network architectures that are trained on the proposed subject split.

References:
[1] Finn, Emily S., et al. “Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity.” Nature neuroscience 18.11 (2015): 1664-1671.
[2] Chen, Shiyang, and Xiaoping Hu. “Individual identification using the functional brain fingerprint detected by the recurrent neural network.” Brain connectivity 8.4 (2018): 197-204.
[3] Chen, Shiyang, and Xiaoping Hu. “Individual identification using the functional brain fingerprint detected by the recurrent neural network.” Brain connectivity 8.4 (2018): 197-204.
[4] Sarar, Gokce, Bhaskar Rao, and Thomas Liu. “Functional connectome fingerprinting using shallow feedforward neural networks.” Proceedings of the National Academy of Sciences 118.15 (2021).
[5] Shojaee, Ali, Kendrick Li, and Gowtham Atluri. “A machine learning framework for accurate functional connectome fingerprinting and an application of a siamese network.” International workshop on connectomics in neuroimaging. Springer, Cham, 2019.
[6] Van Essen, David C., et al. “The Human Connectome Project: a data acquisition perspective.” Neuroimage 62.4 (2012): 2222-2231.
[7] Chicco, Davide. “Siamese neural networks: An overview.” Artificial Neural Networks (2021): 73-94.
[8] Riaz, Atif, et al. “DeepFMRI: End-to-end deep learning for functional connectivity and classification of ADHD using fMRI.” Journal of neuroscience methods 335 (2020): 108506
[9] Ktena, Sofia Ira, et al. “Metric learning with spectral graph convolutions on brain connectivity networks.” NeuroImage 169 (2018): 431-442.