Evaluation of Task fMRI Decoding With Deep Learning on a Small Sample Dataset.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Deep learning (DL) is increasingly used for task-based functional Magnetic Resonance Imaging (fMRI) decoding.
• Challenges include high dimensionality and small sample sizes common in fMRI studies.
• Group-level analyses with limited data suffer from low statistical power and information leakage due to individual variability.

Purpose of the Study:

• To evaluate deep learning models for fMRI decoding under small sample size conditions.
• To assess the impact of data splitting methods on model performance and information leakage.
• To identify the most effective deep learning approach for analyzing bilingual language switching in fMRI data.

Main Methods:

• Utilized a small fMRI dataset focusing on bilingual language switching during a property generation task.
• Compared different deep learning models, including multichannel 2D convolutional neural networks (M2DCNN) and 3D convolutional neural networks (3DCNN).
• Employed controlled data splitting methods, specifically session shuffle split, to mitigate information leakage.

Main Results:

• The multichannel 2D convolutional neural network (M2DCNN) classifier combined with the session shuffle split achieved the highest classification accuracy.
• This approach demonstrated superior performance compared to the 3D convolutional neural network (3DCNN).
• The study highlighted the importance of data folding strategies in managing information leakage.

Conclusions:

• Session shuffle split and M2DCNN provide a robust method for fMRI decoding with limited data.
• The findings suggest that careful consideration of within-subject or within-session information leakage is crucial.
• Further research is needed to fully understand and quantify the complex impact of information leakage in fMRI decoding.