Please don't hesitate to contact me for any questions/details through email: abdullahyr9@gmail.com
More details about the implementation with diagrams and parameters are in the thesis folder. Here my thesis, final presentation, and overleaf zip folder can be found.
input_data - various sets of input data used
output_data - various sets of dictionaries produced and pickled for further processing
Thesis - Folder with thesis and presentation
Dictionary_Testing - Jupyter notebook where extracted dictionaries, sklearn algorithms, PCA and more were tested and plotted
eigenaction_testing - Jupyter notebook where extracted dictionaries from experts and novices were tested and plotted
ESEMD.py - Main algorithm python file
mvnx_video_better.m + mvnx_video_better_original.m - 3D data plotters (incomplete)
nonlinear_dimreduc - Jupyter notebook where nonlinear dimensionality reduction methods were tested
README.md - A highly informative document
requirements.txt - Requirements file
Simply create an environment and run pip install -r requirements.txt
To start an experiment run python ESEMD.py followed by a list of arguments.
For example if you want to run an experiment with parameters of window size 6, output directory experiment_1/w06, data size 50000, the command will look like this: python run_SED.py --window_size 6 --output_name experiment_1/w06 --data_size 50000
More info about the arguments in the 'Arguments' section below
However, input and output files need to be manually changed within the ESEMD.py file if necessary.
ws - window size of ROI, 8 found to be optimal, for eigenactions, their dimensionality will often be as much as the window size due to algorithm
prqt & srqt - reconstruction thresholds for extracted PCs, decreasing them will mean more PCs are considered good and will result in a larger dictionary
st - reconstruction threshold for eigenmotion/eigenaction, increasing this means a higher percentage of the data will need to be accurately reconstructed, resulting in a smaller dictionary
data_size - training data size
dimension_flag - set to: True for eigenmotions, False for eigenactions
Eigenmotion Hyperparameters:
ws - 8
prqt - 0.75
srqt - 0.65
st - 0.01
dimension_flag - True
Eigenaction Hyperparameters:
ws - 8
prqt - 0.95
srqt - 0.85
st - 0.02
dimension_flag - False
Interference Reduction Stage has many values which can be parametrised and further explored, however they currently work well
Understanding surgical actions through signal recognition
Reconstructing signals with eigenmotion analysis
Identification of joint with poorest performance in novice surgeons
