Slice-Wise Augmentation for Multi-Task Learning to Predict Neuropsychological Outcomes After Traumatic Brain Injury

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Traumatic brain injury (TBI) frequently results in persistent cognitive deficits.
• Accurate prediction of these deficits is crucial for tailoring rehabilitation strategies.
• Resting-state functional MRI (fMRI) and clinical data offer potential biomarkers for TBI outcomes.

Purpose of the Study:

• To evaluate the efficacy of multi-task deep learning models in predicting long-term neuropsychological outcomes following TBI.
• To assess the impact of slice-wise data augmentation on model performance using 3D Cerebrovascular reactivity (CVR) maps.

Main Methods:

• Utilized 3D CVR maps from resting-state fMRI acquired 3 months post-TBI.
• Integrated clinical characteristics with CVR maps as input for multi-task deep learning models.
• Compared model performance with and without slice-wise data augmentation for predicting outcomes at 12 months post-TBI.

Main Results:

• The deep learning model incorporating slice-wise augmentation of 3D CVR maps demonstrated superior predictive performance.
• This enhanced model achieved a mean absolute error of 9.18 ± 1.30 in predicting neuropsychological outcomes.
• Slice-wise augmentation significantly improved the accuracy of TBI outcome prediction.

Conclusions:

• Slice-wise augmentation of 3D CVR maps is an effective technique for enhancing deep learning model performance.
• This approach improves the prediction of long-term neuropsychological outcomes in patients with moderate to severe TBI.
• The findings support the use of augmented CVR maps for personalized TBI rehabilitation planning.