Explainable MST-ECoGNet Decode Visual Information from ECoG Signal

TL;DR

This paper introduces MST-ECoGNet, an explainable model that effectively decodes visual information from ECoG signals by leveraging time-frequency, spatial, and real-imaginary features, improving accuracy and model efficiency.

Contribution

The paper proposes a novel explainable model combining mathematical and deep learning methods, revealing how ECoG signals carry visual information and enhancing decoding performance.

Findings

ECoG time-frequency domain contains important visual features.

Spatial features of ECoG signals are crucial for classification.

Combining real and imaginary parts improves decoding accuracy.

Abstract

In the application of brain-computer interface (BCI), we not only need to accurately decode brain signals,but also need to consider the explainability of the decoding process, which is related to the reliability of the model. In the process of designing a decoder or processing brain signals, we need to explain the discovered phenomena in physical or physiological way. An explainable model not only makes the signal processing process clearer and improves reliability, but also allows us to better understand brain activities and facilitate further exploration of the brain. In this paper, we systematically analyze the multi-classification dataset of visual brain signals ECoG, using a simple and highly explainable method to explore the ways in which ECoG carry visual information, then based on these findings, we propose a model called MST-ECoGNet that combines traditional mathematics and deep learning. The main contributions of this paper are: 1) found that ECoG time-frequency domain information carries visual information, provides important features for visual classification tasks. The mathematical method of MST (Modified S Transform) can effectively extract temporal-frequency domain information; 2) The spatial domain of ECoG signals also carries visual information, the unique spatial features are also important features for classification tasks; 3) The real and imaginary information in the time-frequency domain are complementary. The effective combination of the two is more helpful for classification tasks than using amplitude information alone; 4) Finally, compared with previous work, our model is smaller and has higher performance: for the object MonJ, the model size is reduced to 10.82% of base model, the accuracy is improved by 6.63%; for the object MonC, the model size is reduced to 8.78%, the accuracy is improved by 16.63%.