demo_optThresh.m

%% Demo showing how to find best area and probability thresholds from processing 
%% the probability maps of trained data generated by the network

addpath(genpath('Software'))

%% Set names and directories for training data
name = {'501484643','501271265','524691284'};
DirProbMap = ['Results',filesep,'ABO',filesep,'Probability map'];
DirData = ['Dataset',filesep,'ABO'];
DirGTMasks = ['Markings',filesep,'ABO',filesep,'Layer275',filesep,'Grader1'];

%% Setting parameters
pixSize = 0.78;         %um
meanR = 5.85;           %um
AvgArea = round(pi*(meanR/pixSize)^2);
JThresh = 0.5;

% Hyperparameters
probthresh = 0.5:0.05:1;
minArea = [15:5:70];       %pixels:
                           %Should be in ascending order for the current code version 

%% run postprocessinga and performance evaluation for every combination of thresholds
Recall = zeros(numel(name),numel(probthresh),numel(minArea));
Precision = Recall; F1 = Recall;

for i = 1:numel(name)   
    [Recall(i,:,:),Precision(i,:,:),F1(i,:,:)] = multiple_postProcessJaccard([487,487],DirProbMap,DirGTMasks,...
                                                    name{i},AvgArea,minArea,probthresh,JThresh);
end

% Compute mean F1 across train data to find best performance
meanF1 = squeeze(mean(F1,1));
[BestF1,~] = max(meanF1(:));
[indP,indA] = find(meanF1==BestF1,1);

% Thresholds yielding the best performance
minA = minArea(indA);
ProbThresh = probthresh(indP);