SwiFT: Swin 4D fMRI Transformer

Authors: Peter Kim, Junbeom Kwon, Sunghwan Joo, Sangyoon Bae, Donggyu Lee, Yoonho Jung, Shinjae Yoo, Jiook Cha, Taesup Moon

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Swi FT using multiple large-scale resting-state f MRI datasets, including the Human Connectome Project (HCP), Adolescent Brain Cognitive Development (ABCD), and UK Biobank (UKB) datasets, to predict sex, age, and cognitive intelligence. Our experimental outcomes reveal that Swi FT consistently outperforms recent state-of-the-art models.
Researcher Affiliation Academia Peter Yongho Kim Seoul National University Junbeom Kwon Seoul National University Sunghwan Joo Sung Kyun Kwan University Sangyoon Bae Seoul National University Donggyu Lee Sung Kyun Kwan University Yoonho Jung Seoul National University Shinjae Yoo Brookhaven National Lab Jiook Cha Seoul National University Taesup Moon Seoul National University
Pseudocode No The paper includes architectural diagrams but does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Project page: https://github.com/Transconnectome/Swi FT
Open Datasets Yes We evaluate Swi FT using multiple large-scale resting-state f MRI datasets, including the Human Connectome Project (HCP) [24], Adolescent Brain Cognitive Development (ABCD) [25], and UK Biobank (UKB) [26, 27] datasets, to predict sex, age, and cognitive intelligence.
Dataset Splits Yes To evaluate our models, we constructed three random splits with a ratio of (train: validation: test) = (0.7 : 0.15 : 0.15) and reported the average performances across the three splits.
Hardware Specification Yes calculated using a single NVIDIA A100 GPU.
Software Dependencies Yes The major software used for our experiments are as the following: python 3.10.4 pytorch 1.12.1 pytorch-lightning 1.6.5 monai 1.1.0 neptune-client 0.16.4 scipy 1.8.1 torchvision 0.13.1 torchaudio 0.12.1
Experiment Setup Yes For Swi FT, we use the same architecture across all of our experiments, using the architecture corresponding to the Swin Transformer-variant from previous work [19, 23] with a channel number of C = 36. The numbers of layers are fixed to {L1, L2, L3, L4} = {2, 2, 6, 2}... For training, the Binary Cross Entropy (BCE) loss was used for the binary classification task, and the Mean Squared Error (MSE) loss was used for regression tasks.