Table 6.
Comparison of best results presented in this work in comparison with the studies from the state of the art. The three best results for each of the proposed models are presented. It can be observed that the highest accuracy scores were obtained through the single-channel spectrogram images as input for the transfer learning implementations, with scores higher than 90%. Nevertheless, promising results were obtained through the 1D representations of the signal with the 1D CNN models, with accuracies that were able to exceed an 80% accuracy performance. In contrast with the single-channel spectrogram images, these samples required significantly less time to prepare. Additionally, the 9- and 5-channel 2D and 3D representations of the signals provided promising results with the CNN models, with results that were able to outperform the state of the art. Further exploration of these forms of representations as input for DL models might yield promising results for the MI classification task. The best five results are highlighted with bold.
| Model | Input | Accuracy [%] |
|---|---|---|
| [25] | Topographical Map | 65% |
| [26] | 2D Raw EEG signal | 69.82% |
| Single-Layer CNN | 2D Raw 5-Channel EEG Signa | 63.31% |
| 3D Raw 9-Channel EEG Signal | 63.70% | |
| 3D Raw 5-Channel EEG Signal | 62.64% | |
| Two-Layer CNN | 2D Raw 9-Channel EEG Signal | 72.87% |
| 3D Raw 9-Channel EEG Signal | 73.93% | |
| 3D Raw 5-Channel EEG Signal | 69.84% | |
| Three-Layer CNN | Single-Channel Spectrogram Image | 71.08% |
| 3D Raw 9-Channel EEG Signal | 68.35% | |
| 3D Raw 5-Channel EEG Signal | 67.37% | |
|
ResNet18 +
Single-Layer CNN |
Single-Channel Spectrogram Image | 90.55% |
| 9-Channel Vertically Stacked Spectrogram Image |
53.68% | |
| 5-Channel Vertically Stacked Spectrogram Image |
50.19% | |
|
ResNet18 +
Three-Layer CNN |
Single-Channel Spectrogram Image | 93.32% |
| 9-Channel Vertically Stacked Spectrogram Image |
54.70% | |
| 64-Channel Vertically Stacked Spectrogram Image |
50.15% | |
| Single-Layer 1D CNN | 1D Raw EEG Signal Array | 64.69% |
| Two-Layer 1D CNN | 1D Raw EEG Signal Array | 76.29% |
| Three-Layer 1D CNN | 1D Raw EEG Signal Array | 84.87% |
| Four-Layer 1D CNN | 1D Raw EEG Signal Array | 86.12% |
| Five-Layer 1D CNN | 1D Raw EEG Signal Array | 79.58% |