چکیده:
Background. Advances in processing technology have enabled researchers to explore brain mechanisms in greater depth, utilizing methodologies such as deep learning to simulate the electrical activity of the brain with high resolution.
New method. In this study we designed an end-to-end convolutional neural networks (CNNs) model which directly reconstructs electrocorticography signals from image inputs. Image features were extracted specifically for data reconstruction. Additionally, a classification-regressor model was introduced, where a CNN was first trained to classify images into five conceptual categories, and then its extracted features were used with a regressor to reconstruct the brain signals.
Result. Both models were found to be capable of reconstructing ECoG data in the occipital region, which has an important role in the processing of visual information. Furthermore, as the distance from this area increased, the reconstruction accuracy decreased. Another noteworthy finding was that features relevant to the classification of conceptual categories of visual stimuli had more information about the signal, resulting in a greater performance enhancement for regressors (Correlation=0.56, p<0.01) relative to an end-to-end (Correlation=0.46, p<0.05) learning paradigm.
Comparison with existing methods. This paper demonstrated the importance of the feature extraction objective, showing that correctly choosing the model's goal is crucial for enhancing signal reconstruction accuracy.
Conclusions. CNN-based models can effectively simulate the brain network's behavior in generating outputs from input stimulus. When we use image classification to get features, it is better for reconstructing neural signals than when we learn things end to end.
New method. In this study we designed an end-to-end convolutional neural networks (CNNs) model which directly reconstructs electrocorticography signals from image inputs. Image features were extracted specifically for data reconstruction. Additionally, a classification-regressor model was introduced, where a CNN was first trained to classify images into five conceptual categories, and then its extracted features were used with a regressor to reconstruct the brain signals.
Result. Both models were found to be capable of reconstructing ECoG data in the occipital region, which has an important role in the processing of visual information. Furthermore, as the distance from this area increased, the reconstruction accuracy decreased. Another noteworthy finding was that features relevant to the classification of conceptual categories of visual stimuli had more information about the signal, resulting in a greater performance enhancement for regressors (Correlation=0.56, p<0.01) relative to an end-to-end (Correlation=0.46, p<0.05) learning paradigm.
Comparison with existing methods. This paper demonstrated the importance of the feature extraction objective, showing that correctly choosing the model's goal is crucial for enhancing signal reconstruction accuracy.
Conclusions. CNN-based models can effectively simulate the brain network's behavior in generating outputs from input stimulus. When we use image classification to get features, it is better for reconstructing neural signals than when we learn things end to end.
| بازنشر اطلاعات | |
 |
این مقاله تحت شرایط Creative Commons Attribution-NonCommercial 4.0 International License قابل بازنشر است. |
