Journal: Frontiers in Neuroinformatics
Article Title: Robin’s Viewer: Using deep-learning predictions to assist EEG annotation
doi: 10.3389/fninf.2022.1025847
Figure Lengend Snippet: The “Preprocessing” pop-up window is the initial screen of Robin’s Viewer (RV) and has five distinct sections: First, the electroencephalography (EEG)-file selection (see Section “3.1 Loading EEG data”), where the user can load their EEG recording and display selected statistics about the data once it is loaded. Second, the preprocessing settings (see Section “3.2 Preprocessing and visualization settings”), which are used to set a bandpass filter [finite impulse response (FIR) filter with Blackman window] and a custom reference. Third, the bad-channel handling (see Sections “3.2 Preprocessing and visualization settings” and “3.4 Bad-channel marking”), where the user can decide whether to use automatic bad channel detection, and whether to interpolate bad channels. Fourth, the visualization settings (see Section “3.2 Preprocessing and visualization settings”), comprised of downsampling, custom scaling (by default 1e-6 as RV scales data from volts to microvolts for plotting), the gap between traces (by default 40 (μV); setting this to 0 results in butterfly mode where all traces are collapsed on top of each other; values higher than 40 move traces further apart), segment length to plot [by default 60 (seconds)], whether or not to activate the view-slider, and selection of channels to plot. Visualization settings will only be applied to the data for plotting and hence will not be saved in the save-file (in contrast to the preprocessing settings). Fifth, the deep-learning model settings (see Sections “3.2 Preprocessing and visualization settings” and 3.5 Deep-learning model predictions”), where previously saved model output can be loaded and the integrated deep-learning model can be activated to generate predictions. Clicking the “Plot” button at the bottom will close this window and, after a loading screen (which lasts as long as it takes to plot the data), it will open up the main graphical user interface (GUI).
Article Snippet: All plots and visualizations, as well as parts of the interactive GUI, including the slider and artifact annotation mechanism, included in RV, were written using Plotly.
Techniques: Selection
![The “Preprocessing” pop-up window is the initial screen of Robin’s Viewer (RV) and has five distinct sections: First, the electroencephalography (EEG)-file selection (see Section “3.1 Loading EEG data”), where the user can load their EEG recording and display selected statistics about the data once it is loaded. Second, the preprocessing settings (see Section “3.2 Preprocessing and visualization settings”), which are used to set a bandpass filter [finite impulse response (FIR) filter with Blackman window] and a custom reference. Third, the bad-channel handling (see Sections “3.2 Preprocessing and visualization settings” and “3.4 Bad-channel marking”), where the user can decide whether to use automatic bad channel detection, and whether to interpolate bad channels. Fourth, the visualization settings (see Section “3.2 Preprocessing and visualization settings”), comprised of downsampling, custom scaling (by default 1e-6 as RV scales data from volts to microvolts for plotting), the gap between traces (by default 40 (μV); setting this to 0 results in butterfly mode where all traces are collapsed on top of each other; values higher than 40 move traces further apart), segment length to plot [by default 60 (seconds)], whether or not to activate the view-slider, and selection of channels to plot. Visualization settings will only be applied to the data for plotting and hence will not be saved in the save-file (in contrast to the preprocessing settings). Fifth, the deep-learning model settings (see Sections “3.2 Preprocessing and visualization settings” and 3.5 Deep-learning model predictions”), where previously saved model output can be loaded and the integrated deep-learning model can be activated to generate predictions. Clicking the “Plot” button at the bottom will close this window and, after a loading screen (which lasts as long as it takes to plot the data), it will open up the main <t>graphical</t> <t>user</t> <t>interface</t> <t>(GUI).</t>](https://pub-med-central-images-cdn.bioz.com/pub_med_central_ids_ending_with_1202/pmc09951202/pmc09951202__fninf-16-1025847-g001.jpg)
![(A) The main graphical user interface (GUI) of Robin’s Viewer (RV) with its six subsections highlighted [see Section “3.3 Graphical user interface (GUI)”]: The menu bar ( red ) contains buttons to open pop-up windows for file loading and preprocessing, saving, annotation statistics, power spectrum, help, and shutting down RV, as well as a button to recalculate the deep-learning model’s predictions (if activated) and left and right arrow buttons used to navigate the currently viewed timeframe (if the data is loaded in segments). The taskbar ( light blue ) seen in detail in panel (B) . The labeled buttons ( dark blue ): The first two buttons are used to reset the view, either across the channel- (“Reset channel-axis”) or time-axis (“Reset time-axis”). The third button (“Hide/show bad channels”) allows to hide marked bad channels from the plot. The fourth button (“Highlight model-channels”) only appears when deep-learning predictions are activated and highlights the channels used by the model to make its predictions in blue. The view-slider ( green ) is used to scroll continuously along the time-axis of the signal. The legend ( purple ) shows all available channels. Clicking on any channel in the legend once will hide it from the plot (clicking again reverses this). Double-clicking on a channel will hide all other channels from the plot, except for the selected channel. The plot ( yellow ) shows the traces for all selected channels spread across the vertical axis, and time (in seconds) on the horizontal axis. The user can hover over any given point in the plot in order to display the time (in seconds) and amplitude (in μV if no custom scaling was used) values of the trace under the mouse. The deep-learning predictions, if activated (see Section “3.5 Deep-learning model predictions”), are plotted below the EEG traces. (B) The taskbar hosts ten buttons from left to right: (1) Take a picture of the (annotated) electroencephalography (EEG) signal and download it as a .png file. (2) Select an area to zoom in on. (3) Move view along the channel- or time-axis. (4) Select a channel to mark it as “bad” or select a segment of the data for which to calculate and display the main underlying frequency and power spectrum. (5) Select a segment of the plot to annotate. (6) Delete the currently selected annotation. (7) Zoom in one step. (8) Zoom out one step. (9) Zoom out as much as necessary to show all channels for the entire duration of the recording (or segment). (10) Display a ruler from both axes to the datapoint currently hovered on.](https://pub-med-central-images-cdn.bioz.com/pub_med_central_ids_ending_with_1202/pmc09951202/pmc09951202__fninf-16-1025847-g002.jpg)
