calculateDecodePerformance.m (291ee33) - SVMCrossVal.git

calculateDecodePerformance.m

% function [decodePerformance rawTimecourse ] = calculateDecodePerformance(des,timeLineStart, timeLineEnd, decodeDuration, svmargs, conditionList, sessionList, voxelList, classList, labelMap,normalize)
function outputStruct = calculateDecodePerformance(inputStruct)

addpath 'libsvm-mat-2.88-1';

outputStruct = struct;

des             = inputStruct.des;
timeLineStart   = inputStruct.frameShiftStart;
timeLineEnd     = inputStruct.frameShiftEnd;
decodeDuration  = inputStruct.decodeDuration;
svmargs         = inputStruct.svmargs;
sessionList     = inputStruct.sessionList;
voxelList       = inputStruct.voxelList;
% classList       = inputStruct.classList;
% labelMap        = inputStruct.labelMap;
smoothed       = inputStruct.smoothed;
globalStart     = inputStruct.psthStart;
globalEnd       = inputStruct.psthEnd;
baselineStart   = inputStruct.baselineStart;
baselineEnd     = inputStruct.baselineEnd;
eventList       = inputStruct.eventList;

minPerformance = inf;
maxPerformance = -inf;

%Pro Voxel PSTH TIMELINE berechnen.
        %   timeshift mit pst-timeline durchf�hren.
        % psth-timeline -25 bis +15 zu RES Onset.
        
%         eventList       = [9,11,13;10,12,14];
%         globalStart     = -25;
%         globalEnd       = 15;
%         baselineStart   = -22;
%         baselineEnd     = -20;
        
        
        for voxel = 1:size(voxelList,1)  % [[x;x],[y;y],[z;z]]
                extr  = calculateImageData(voxelList(voxel,:),des,smoothed);
                rawdata=cell2mat({extr.mean}); % Raw Data
                pst{voxel}  = calculatePST(des,globalStart,baselineStart,baselineEnd,globalEnd,eventList,rawdata,sessionList);
        end

decodePerformance = [];

for timeShift   = timeLineStart:1:timeLineEnd
            frameStart  = floor(-globalStart+1+timeShift - 0.5*decodeDuration);
            frameEnd    = min(ceil(frameStart+decodeDuration + 0.5*decodeDuration),-globalStart+globalEnd);
            
            tmp =[];
            anyvoxel = 1;
            for label = 1:size(pst{1,anyvoxel},2) 
                for dp = 1:size(pst{1,anyvoxel}{1,label},1) % data point
                row = label;
                    for voxel = 1:size(pst,2)
                        row = [row, pst{1,voxel}{1,label}(dp,frameStart:frameEnd)]; % label,value,value
                    end
                tmp  = [tmp; row];
                end
            end 
        
            svmdata      = tmp(:,2:size(tmp,2));
            svmlabel     = tmp(:,1);
            performance  = svmtrain(svmlabel, svmdata, svmargs);

minPerformance = min(minPerformance,performance);
            maxPerformance = max(maxPerformance,performance);

decodePerformance = [decodePerformance; performance];
        end
        
        outputStruct.decodePerformance  = decodePerformance;
        outputStruct.svmdata            = svmdata;
        outputStruct.svmlabel           = svmlabel;
        outputStruct.rawTimeCourse      = pst;
        outputStruct.minPerformance     = minPerformance;
        outputStruct.maxPerformance     = maxPerformance;

% display(sprintf('Min CrossVal Accuracy: %g%% \t Max CrossVal Accuracy: %g%%',minPerformance,maxPerformance));
end

function extr = calculateImageData(voxelList,des,smoothed)

dtype='PSTH';

switch dtype 
    case 'PSTH'
        V=des.xY.VY;
    case 'betas'
        V=des.Vbeta;
end;

if (~smoothed)
  for z=1:length(V) % Change Drive Letter!\
      % D:....SUBJECTID\session\swfanders...
      % D:....SUBJECTID\session\wfanders...
      tmp = findstr(filesep,V(z).fname);
      V(z).fname = strcat(V(z).fname(1:tmp(end)),V(z).fname(tmp(end)+2:end));
  end;
end

% rad = 0; % one voxel
% opt = 1; % xyz coordinates [mm]

vox = voxelList;
nRoi = size(vox,1);

nImg = numel(V);

for k=1:nImg
	extr(k) = struct(...
        'val',   repmat(NaN, [1 nRoi]),...
		'mean',  repmat(NaN, [1 nRoi]),...
		'sum',   repmat(NaN, [1 nRoi]),...
		'nvx',   repmat(NaN, [1 nRoi]),...
		'posmm', repmat(NaN, [3 nRoi]),...
		'posvx', repmat(NaN, [3 nRoi]));

roicenter = round(inv(V(k).mat)*[vox, ones(nRoi,1)]');

for l = 1:nRoi

%         if rad==0
            x = roicenter(1,l);
            y = roicenter(2,l);
            z = roicenter(3,l);
%         else
%             tmp = spm_imatrix(V(k).mat);
%             vdim = tmp(7:9);
%             vxrad = ceil((rad*ones(1,3))./(ones(nRoi,1)*vdim))';
%             [x y z] = ndgrid(-vxrad(1,l):sign(vdim(1)):vxrad(1,l), ...
%                       -vxrad(2,l):sign(vdim(2)):vxrad(2,l), ...
%                       -vxrad(3,l):sign(vdim(3)):vxrad(3,l));
%             sel = (x./vxrad(1,l)).^2 + (y./vxrad(2,l)).^2 + ...
%                   (z./vxrad(3,l)).^2 <= 1;
%             x = roicenter(1,l)+x(sel(:));
%             y = roicenter(2,l)+y(sel(:));
%             z = roicenter(3,l)+z(sel(:));
%         end;
		dat                 = spm_sample_vol(V(k), x, y, z,0);
		[maxv maxi]         = max(dat);
		tmp                 = V(k).mat*[x(maxi); y(maxi); z(maxi);1]; % Max Pos
		extr(k).val(l)      = maxv;
		extr(k).sum(l)      = sum(dat);
		extr(k).mean(l)     = nanmean(dat);
        extr(k).nvx(l)      = numel(dat);
		extr(k).posmm(:,l)  = tmp(1:3);
		extr(k).posvx(:,l)  = [x(maxi); y(maxi); z(maxi)]; % Max Pos
	end;

end;
end

% disp(sprintf('Extracted at %.1f %.1f %.1f [xyz(mm)], average of %i voxel(s) [%.1fmm radius Sphere]',vox,length(x),rad));

function pst = calculatePST(des,globalStart,baselineStart,baselineEnd,globalEnd,eventList,data,sessionList)
    bstart          = baselineStart;
    bend            = baselineEnd;
    edur            = 12;
    pre             =  globalStart;
    post            =  globalEnd;
    res             = 1;

normz           = 'file';
    pm              = 0;

lsess           = getNumberOfScans(des);
    nSessions       = getNumberOfSessions(des);
    tr              = 2;

[evntrow evntcol]=size(eventList);

hsec=str2num(des.xsDes.High_pass_Filter(8:end-3)); % Highpass filter [sec] from SPM.mat

if strcmp(des.xBF.UNITS,'secs')
        unitsecs=1;
    end;

nScansPerSession=getNumberOfScans(des);
    %stime=[0:tr:max(nScansPerSession)*tr+post-tr]; % Stimulus time for raw data plot
    stime=0:tr:max(nScansPerSession)*tr+round(post/tr)*tr-tr; % Stimulus time for raw data plot

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
    % RUN
    %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Digital Highpass
    Rp=0.5;
    Rs=20;
    NO=1;
    Wp=1/((1/2/tr)/(1/hsec));
    [B, A] = ellip(NO,Rp,Rs,Wp,'high');

sdata(1:max(nScansPerSession)+round(post/tr),1:nSessions)=nan; % Open Data Matrix
    for z=1:nSessions % Fill Data Matrix sessionwise
        sdata(1:nScansPerSession(z),z)=data(sum(nScansPerSession(1:z))-nScansPerSession(z)+1:sum(nScansPerSession(1:z)))';
    end;
%         usdata=sdata; % Keep unfiltered data

sdatamean=nanmean(nanmean(sdata(:,:)));
    for z=1:nSessions
%             X(:,z)=[1:1:max(nScansPerSession)]'; % #Volume
        sdata(1:nScansPerSession(z),z)=filtfilt(B,A,sdata(1:nScansPerSession(z),z)); %Filter Data (Highpass)
    end;
    sdata=sdata+sdatamean;

%%%%Parametric Modulation Modus%%%%
    if pm %Find Parameters for Event of Interest
        [imods modss mods erow evntrow eventList] = getParametricMappingEvents(eventList,evntrow,des,pmf);
    end;
    %%%%PM%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

for zr=1:evntrow
        n{zr}=0;
        nn{zr}=0; 
        nnn{zr}=0;
        sstart{zr}=1;
    end;

sesst0=0; 
    for sessionID=sessionList
        if sessionID>1
            sesst0(sessionID)=sum(lsess(1:sessionID-1))*tr;  
        end;
        for zr=1:evntrow  %LABEL NUMBER, EVENT GROUP
            sstart{zr}=n{zr}+1;
            for ze=1:evntcol % EVENT INDEX in EventList
                if ze==1 || (ze>1 && eventList(zr,ze)~=eventList(zr,ze-1))
                    for zz=1:length(des.Sess(sessionID).U(eventList(zr,ze)).ons) % EVENT REPETITION NUMBER
                        if ~unitsecs
                            des.Sess(sessionID).U(eventList(zr,ze)).ons(zz)=(des.Sess(sessionID).U(eventList(zr,ze)).ons(zz)-1)*tr;
                            des.Sess(sessionID).U(eventList(zr,ze)).dur(zz)=(des.Sess(sessionID).U(eventList(zr,ze)).dur(zz)-1)*tr;
                        end;

nnn{zr}=nnn{zr}+1; % INFO for rawdataplot start
                        if des.Sess(sessionID).U(eventList(zr,ze)).dur(zz)<edur
                            mev{zr}(nnn{zr},1:2)=[des.Sess(sessionID).U(eventList(zr,ze)).ons(zz)+sesst0(sessionID) edur]; % modeled event [onset length]
                        else
                            mev{zr}(nnn{zr},1:2)=[des.Sess(sessionID).U(eventList(zr,ze)).ons(zz)+sesst0(sessionID) des.Sess(sessionID).U(eventList(zr,ze)).dur(zz)];
                        end; % INFO for rawdataplot end

n{zr}=n{zr}+1;
                        pst{zr}(n{zr},:)=interp1(stime,sdata(:,sessionID),[des.Sess(sessionID).U(eventList(zr,ze)).ons(zz)+pre:res:des.Sess(sessionID).U(eventList(zr,ze)).ons(zz)+post],'linear');
                        if strcmp(normz,'epoc')
                            bline=nanmean(pst{zr}(n{zr},round(-pre/res+(bstart)/res+1):round(-pre/res+(bend)/res+1)));
                            if isnan(bline)
                                pst{zr}(n{zr},1:-pre/res+post/res+1)=nan;
                            else
%                                     nn{zr}=nn{zr}+1;
                                pst{zr}(n{zr},:)=(pst{zr}(n{zr},:)-bline)/bline*100; % 'epoch-based' normalization
                            end;
                        end;
                    end;
                end;
            end;
            if ~strcmp(normz,'epoc')
                bline(zr)=nanmean(nanmean(pst{zr}(sstart{zr}:n{zr},-pre/res+(bstart)/res+1:-pre/res+(bend)/res+1)));
                bstd(zr)=nanmean(nanstd(pst{zr}(sstart{zr}:n{zr},-pre/res+(bstart)/res+1:-pre/res+(bend)/res+1)));
                nn{zr}=n{zr};
            end;
        end;
        if strcmp(normz,'filz')
            for zr=1:evntrow
                pst{zr}(sstart{zr}:n{zr},:)=(pst{zr}(sstart{zr}:n{zr},:)-mean(bline))/mean(bstd); % session-based z-score normalization
            end;
        elseif strcmp(normz,'file')
            for zr=1:evntrow
                pst{zr}(sstart{zr}:n{zr},:)=(pst{zr}(sstart{zr}:n{zr},:)-mean(bline))/mean(bline)*100; % session-based normalization
            end;
        end;
    end;
end

SVMCrossVal.git

killed unnecessary assignin raw axis fixed value added smooth option