Requirment:

1. tensorflow 2.2.0
2. dipy
3. nibabel
4. fury
5. plyfile

Running tip: create a tensorflow-cpu environment to run is safer since you might have a memory problem if you are using tensorflow-gpu and your gpu memory is not high.

python Detection.py -f 156031_ex_cc-body_shore.ply

the terminal will print the following and generate the results
model_type = 0,default choice
file_name = '156031_ex_cc-body_shore.ply'
threshold control = 80, default value
length control = 40, default value

If you run this sample, you could find two files in the folder /result
-156031_ex_cc-body_shore_cleaned_m0.ply
the anomaly removed version
-Detection156031_m0.npy
a list of the detection result 1：normal 0：anomaly

• -m Specify the model
  Default: 0. the model choice will be explained later
• -f Input the file name
  the file is in /data folder, there is an sample
• -t Specify the threshold percentile
  Default: 85. The Autoencoder reconstruct the input fibers and we calculate the reconstruction MSE, rank the MSE and set a percentile as the detection threshold.
• -l Specify length under what value is anomalous
  Default: 40.

Detection_demo.ipynb offers two methods to visualize the result:

1. a binary visualization of the detection
   red: normal curves
   white : anomalous curves
2. anomalous fibers removal

You can choose one of the following models
models 0，1，2，3，6，9，10 has better result.
The implementation of these models could be found in /models.

index: 0       model: deep Bi GRU
index: 1       model: deep Bi LSTM
index: 2       model: deep GRU
index: 3       model: deep LSTM
index: 4       model: Bi GRU
index: 5       model: Undercomplete GRU
index: 6       model: Autoencoder GRU
index: 7       model: Bidirectional LSTM
index: 8       model: Undercomplete LSTM
index: 9       model: Autoencoder LSTM
index: 10      model: Autoencoder RNN
index: 11      model: Bidirectional RNN
index: 12      model: Undercomplete RNN

The Seq2seq related models are implemented in a different way, so we create a new folder /Seq_models to store the weights and code. the **-Detection.ipynb demonstrated how to generate the detection result.

In performance_analysis.ipynb, we demonstrated how to evaluate the Anomaly detection performance between different models and with the manually Detection results.