SAD: semi-supervised automatic detection of BOLD activations in high temporal resolution fMRI data

Area of Science:

• Neuroimaging
• Functional Magnetic Resonance Imaging (fMRI) analysis

Background:

• The general linear model (GLM) is widely used in fMRI analysis but relies on a pre-defined hemodynamic response function (HRF), which can reduce reliability and distort inferences.
• A fixed HRF assumption may not accurately represent individual or voxel-specific hemodynamic responses, limiting the precision of fMRI studies.

Purpose of the Study:

• To introduce a novel semi-supervised automatic detection (SAD) method for fMRI data analysis.
• To overcome the limitations of the GLM by eliminating the need to pre-specify the HRF shape.
• To enhance the reliability and accuracy of fMRI data interpretation.

Main Methods:

• The SAD method utilizes a Bi-LSTM neural network for classifying high temporal resolution fMRI data (75-ms).
• Network training was performed iteratively on an fMRI dataset with 75-ms temporal resolution.
• Performance was validated on a separate fMRI dataset from the same participant and compared against the standard GLM approach.

Main Results:

• The SAD method demonstrated strong classification performance with a true-positive rate of 0.961 and an area under the receiver operating curve of 0.998.
• Achieved a true-negative rate of 0.99, F1-score of 0.979, and very low false discovery and positive rates (0.002).
• These results were obtained at a high temporal resolution of 75 ms.

Conclusions:

• The SAD method successfully detects hemodynamic responses at 75-ms temporal resolution without requiring a specific HRF model.
• This approach offers a more flexible and potentially more reliable alternative to the GLM for fMRI analysis.
• Future research should explore SAD's applicability across more participants and diverse fMRI experimental paradigms.