Generalizable Representation Learning for fMRI-based Neurological Disorder Identification
Authors: Wenhui Cui, Haleh Akrami, Anand Joshi, Richard Leahy
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Results demonstrate the superiority of our representation learning strategy on diverse clinically-relevant tasks. We evaluate the Me TSK model pre-trained with a source and a target dataset on unseen clinical datasets. Table 1: A comparison of mean AUCs of 5-fold cross-validation on ADHD dataset and ABIDE dataset using different methods: fine-tuning, multi-task learning, the proposed strategy Me TSK, and other baseline methods. |
| Researcher Affiliation | Academia | Wenhui Cui EMAIL Ming Hsieh Department of Electrical and Computer Engineering University of Southern California |
| Pseudocode | No | The paper describes the bi-level optimization strategy with detailed steps (e.g., 'Outer loop (M iterations): Step 1. Initialize the target head...', 'Step 2. Inner loop (k update steps): Only target head parameters θt are updated...'), but it does not present this information in a structured pseudocode block or an explicitly labeled algorithm. |
| Open Source Code | No | The paper mentions external pre-trained models for comparison: 'Their pre-trained model is publicly available at https://github.com/athms/learning-from-brains.' and 'Similarly, we directly applied their pre-trained model available at https://github.com/vandijklab/Brain LM to generate features.' However, there is no explicit statement or link provided for the source code of the authors' own proposed methodology (Me TSK). |
| Open Datasets | Yes | We use HCP (Van Essen et al., 2013) as our source dataset due to its large size. For target datasets, we use the ADHD (Bellec et al., 2017) datasets, and the ABIDE dataset (Craddock et al., 2013) during the training of the Me TSK model. We use the preprocessed data released on (http: //preprocessed-connectomes-project.org/adhd200/). The Autism Brain Imaging Data Exchange I (ABIDE I) (Craddock et al., 2013)... We downloaded the data from http://preprocessed-connectomes-project. org/abide/. The f MRI data are available to download from FITBIR (https://fitbir. nih.gov). Open Access Series of Imaging Studies (OASIS-3) (La Montagne et al., 2019) is a longitudinal neuroimaging... and is publicly available at https://www.oasis-brains.org. Both the Neurocon and Tao WU datasets were downloaded from https://fcon_1000.projects.nitrc.org/indi/retro/parkinsons.html. |
| Dataset Splits | Yes | We use 5-fold cross-validation to split training/testing sets on ADHD-Peking/ABIDE-UM data and use all HCP data for training. For meta-learning, the target training set in each fold is further divided into a meta-training set XTtr of 157 subjects and a meta-validation set XTval of 39 subjects. Both the meta-train and meta-validation sets are randomly re-sampled in every iteration... |
| Hardware Specification | Yes | Training Me TSK takes approximately 6 hours on a single NVIDIA V100 GPU with 32 GB of memory. |
| Software Dependencies | No | We use an Adam optimizer (Kingma & Ba, 2014) with learning rate β = 0.001 in the outer loop, and an SGD optimizer (Ketkar, 2017) with learning rate α = 0.01 in the inner loop. While specific optimizers are mentioned, no version numbers for programming languages (e.g., Python) or software libraries (e.g., PyTorch, TensorFlow) are provided. |
| Experiment Setup | Yes | The length of input sub-sequences for ST-GCN is fixed at 128. We generate one meta-training batch by randomly selecting an equal number of samples from each class. The batch size is 32, both for the meta-training and the meta-validation set. We use an Adam optimizer (Kingma & Ba, 2014) with learning rate β = 0.001 in the outer loop, and an SGD optimizer (Ketkar, 2017) with learning rate α = 0.01 in the inner loop. The number of inner loop update steps is 25. We set the hyper-parameter λ = 30 and the temperature parameter τ = 30 to adjust the scale of losses... |