Oracle_LOSO.wiki
We are now ready to test NeuroGPS using our dataset.
The test file is F1_ORACLE_LOSO.py, where are implemented the Oracle and Leave-One-Stack-Out functions, described in Tests Overview section. To start the computation it's sufficient to run the F1_ORACLE_LOSO.py, all it's already setted.
The computation is divided into two parts:
- Oracle function. For each stack, it's calculated the couple of treshold-minimum radius that realize the best F1-score. All the results are desplayed. At the best treshold, minimum radius and F1-score are hghlightd for each stack.
Treshold and min. radius: T7_R5 DEBUG reading treshold\treshold_7\T7_R5\x007038_T7_R5.marker... Solving max weight matching (840 nodes, 544 edges) |pred|=189 |true|=651 P: 93.65 / R: 27.19 / F1: 42.14 ==== TP: 177 / FP: 12 / FN: 474
A typical output that describes the F1-score calculation for the output of NeuroGPS for the stack x007038 with treshold = 7 and minimum radius = 5
|SUBSTACK|: x004647.marker |TRESHOLD AND MIN RADIUS USED|: T4_R3 |BEST F1|: 78.86
A typical final output ot the Oracle function, that indicates the treshold and minimum radius values that realizes the best F1-score for the x004647 stack.
- Leave-One-Stack-Out function. LOSO function is implemented in the second part of the code and uses the results obtained by the Oracle function. LOSO indicate to use the treshold and minimum radius values, which allowed to realize the higher F1-score among all the other n-1 stack of images, for the nth stack.
LOSO FOR THE STACK: x013442.marker . CALCULATING F1 WITH TRESHOLD AND MIN RADIUS T5_R5 Solving max weight matching (257 nodes, 141 edges) |pred|=116 |true|=141 P: 92.24 / R: 75.89 / F1: 83.27 ==== TP: 107 / FP: 9 / FN: 34
A typical LOSO output that indicates the F1 result for the stack x013443 using the predicted parametres treshold = 5 and minimum radius = 5.
The aim using the LOSO function it's to find a set of parametres treshold_minimum radius that compute a good F1-score for all stacks.
Legend for read the output:
- pred indicates the neurons individuated by NeuroGPS;
- true indicates the neurons actually in the analized stack;
- P indicates the precision;
- R indicates the recall;
- TP indicates the true positivs;
- FP indicates the false positives;
- FN indicates the false negatives;
A graph that indicates how pecision, recall and F1 perform whitin NeuroGPS respect to our dataset.