​ To promote the development of this field, particularly to advance interpretability techniques in neuroimaging, we have decided to open-source our dynamic visualization solution.

• Input: Attention results from multiple epochs

• Function:

  • K-fold cross validation averages can be calculated and displayed separately, and exported to PDF and PNG for display.

  • Provide the ranking of important brain areas in text form.

  • Provide a custom parameter interface.

    

  • Three types of diagrams (Here's the demo on the SZ dataset)

    • Average attention heatmap showing the evolution of attention weights across training epochs and independent components.
    
    • Temporal dynamics of the top 5 most important IC components throughout training epochs. Lines show mean attention weights with shaded areas representing standard deviation across folds, highlighting the stability and convergence patterns of critical brain regions
    
    • Standard deviation heatmap illustrating the variability of attention weights across different folds
    

• GPU: NVIDIA GeForce RTX 2080Ti (11GB VRAM)
• CUDA: 10.0
• cuDNN: 7.4
• GPU Driver: 410.48

• Python: 3.7
• Operating System: Ubuntu 18.04 LTS
• TensorFlow: 1.13.1 with Keras 2.2.4

conda create -n cfml
conda activate cfml
pip install -r requirements.txt

We use Keras for training and sklearn.metrics for evaluation. The code execution command is as follows:

python main_tc_fnc.py