Tests.wiki

Now that we have the images stacks in 3D thanks to Fiji (see the Fiji section) and the NeuroGPS software is initialized (see the NeuroGPS section) we can proceed with the tests.

The tests are used to verify the accuracy of the algorithm in detecting neurons. To do this we use the eval_perf.py code, where is implemented the function that computes the precision, the recall and the F1-score of a NeuroGPS's output.

The main ideas behind the test are:

1. to find the couple of values treshold-minimum radius for each 3D image in the dataset that allows to compute the best F1-score. This is done by the Oracle function.
2. to predict a couple of values treshold-minimum radius for each 3D image in the dataset that allows to compute a good F1-score, based on the treshold and minimum radius parametres used to compute the best F1-score for the other images in the dataset. This is done by the Leave-One-Stack-Out function.

For use the Oracle and LOSO function we need to obtain the outputs of NeuroGPS for each substantial combination of treshold and minimum radius for each stack of images. All these files have been computed in batch with NeuroGPS and saved in the treshold folder.

We have to adapt the NeuroGPS outputs to the GT marker files to proceed with the tests. See Adapt the NeuroGPS outputs to the GT for more details.

Once adapted the outputs, the task is:

• calulate all the F1-score for each output for each 3D image and find the output that realizes the best one (ORACLE)
• calculate the F1-score for the output obtained with the parametres predicted by the LOSO function.

To compute these F1-scores we use the F1_ORACLE_LOSO.py comparing each output with the corresponding GT marker file.