项目笔记《DeepLung:Deep 3D Dual Path Nets for Automated Pulmonary Nodule Detection and Classification》（三）（下）结果评估
在（上）中讲了如何得到csv文件并调用noduleCADEvaluationLUNA16.py求取froc值，这里就讲一讲froc值是如何求取的。
annotations_filename = './annotations/annotations.csv' annotations_excluded_filename = './annotations/annotations_excluded.csv' seriesuids_filename = './annotations/seriesuids.csv' results_filename = './annotations/3DRes18FasterR-CNN.csv'#3D Faster R-CNN - Res18.csv' #top5.csv'# seriesuids_path = '/home/ustc/lclin/code/DeepLung/evaluationScript/series_uid9.csv' results_path = "/home/ustc/lclin/code/DeepLung/detector/results/wzy_res18/retrft969/val120/predanno-0.125pbb.csv" noduleCADEvaluation(annotations_filename,annotations_excluded_filename,seriesuids_path,results_path,'./')
如上面代码所示，输入标签文件，结果文件，调用noduleCADEvaluation函数即可，任务完成，十分简单。
接下来看一下这个函数。
def noduleCADEvaluation(annotations_filename,annotations_excluded_filename,seriesuids_filename,results_filename,outputDir): ''' function to load annotations and evaluate a CAD algorithm @param annotations_filename: list of annotations @param annotations_excluded_filename: list of annotations that are excluded from analysis @param seriesuids_filename: list of CT images in seriesuids @param results_filename: list of CAD marks with probabilities @param outputDir: output directory ''' print annotations_filename (allNodules, seriesUIDs) = collect(annotations_filename, annotations_excluded_filename, seriesuids_filename) #根据标签和用户id，求出所有结节，所有用户 evaluateCAD(seriesUIDs, results_filename, outputDir, allNodules, #根据结节，用户，结果文件，输出froc值 os.path.splitext(os.path.basename(results_filename))[0], maxNumberOfCADMarks=100, performBootstrapping=bPerformBootstrapping, numberOfBootstrapSamples=bNumberOfBootstrapSamples, confidence=bConfidence)
这个函数调用了两个函数collect和evaluateCAD。先从collect开始看。
def collect(annotations_filename,annotations_excluded_filename,seriesuids_filename): annotations = csvTools.readCSV(annotations_filename) annotations_excluded = csvTools.readCSV(annotations_excluded_filename) seriesUIDs_csv = csvTools.readCSV(seriesuids_filename) seriesUIDs = [] #建立一个用户列表，将用户id添加进去 for seriesUID in seriesUIDs_csv: seriesUIDs.append(seriesUID[0]) #seriesUID也是一个列表，只不过只有一个元素 allNodules = collectNoduleAnnotations(annotations, annotations_excluded, seriesUIDs) return (allNodules, seriesUIDs)
该函数又调用了另一个函数collectNoduleAnnotations。
def collectNoduleAnnotations(annotations, annotations_excluded, seriesUIDs): allNodules = {} #将所有结节存储在字典中 noduleCount = 0 noduleCountTotal = 0 for seriesuid in seriesUIDs: # print 'adding nodule annotations: ' + seriesuid nodules = [] numberOfIncludedNodules = 0 #对真正用来检测的结节计数 # add included findings header = annotations[0] for annotation in annotations[1:]: nodule_seriesuid = annotation[header.index(seriesuid_label)] if seriesuid == nodule_seriesuid: #将结节所属的用户id与要建立字典索引的id比较，若相同，就获取它 nodule = getNodule(annotation, header, state = "Included") nodules.append(nodule) numberOfIncludedNodules += 1 # add excluded findings header = annotations_excluded[0] for annotation in annotations_excluded[1:]: nodule_seriesuid = annotation[header.index(seriesuid_label)] if seriesuid == nodule_seriesuid: #对无关结节也执行上面的操作，不同的是不对include计数 nodule = getNodule(annotation, header, state = "Excluded") nodules.append(nodule) allNodules[seriesuid] = nodules noduleCount += numberOfIncludedNodules noduleCountTotal += len(nodules) print 'Total number of included nodule annotations: ' + str(noduleCount) print 'Total number of nodule annotations: ' + str(noduleCountTotal) return allNodules
这段代码比较简单，其中有一个获取结节的函数getNodule，需要看一下
def getNodule(annotation, header, state = ""): nodule = NoduleFinding() nodule.coordX = annotation[header.index(coordX_label)] #依次将x,y,z添加到nodule对象的属性中 nodule.coordY = annotation[header.index(coordY_label)] nodule.coordZ = annotation[header.index(coordZ_label)] if diameter_mm_label in header: #检查有无直径标签，有则添加 nodule.diameter_mm = annotation[header.index(diameter_mm_label)] if CADProbability_label in header: #检查有无概率标签，有则添加 nodule.CADprobability = annotation[header.index(CADProbability_label)] if not state == "": nodule.state = state return nodule
这里又又出现一个函数，准确地说是类，NoduleFinding()，这个类存在于nodulefinding.py模块，也是在evaluationScript这个文件夹中。
一起来欣赏下这个类，一目了然，而且这个模块也只有这段代码。
class NoduleFinding(object): ''' Represents a nodule ''' def __init__(self, noduleid=None, coordX=None, coordY=None, coordZ=None, coordType="World", CADprobability=None, noduleType=None, diameter=None, state=None, seriesInstanceUID=None): # set the variables and convert them to the correct type self.id = noduleid self.coordX = coordX self.coordY = coordY self.coordZ = coordZ self.coordType = coordType self.CADprobability = CADprobability self.noduleType = noduleType self.diameter_mm = diameter self.state = state self.candidateID = None self.seriesuid = seriesInstanceUID
另外，大家可能会奇怪，xxx_label都是些什么，这些在文件noduleCADEvaluationLUNA16.py中都有定义，如下，看一下标签文件就明白了。
seriesuid_label = 'seriesuid' coordX_label = 'coordX' coordY_label = 'coordY' coordZ_label = 'coordZ' diameter_mm_label = 'diameter_mm' CADProbability_label = 'probability'
最后，回到起点，看看froc究竟是如何计算的，有请evaluateCAD函数，这段代码超级长。
def evaluateCAD(seriesUIDs, results_filename, outputDir, allNodules, CADSystemName, maxNumberOfCADMarks=-1,#输入病人id，检测结果，评估结果的输出目录，所有标签结节， performBootstrapping=True,numberOfBootstrapSamples=1000,confidence = 0.95): #检测系统的名字，其余参数与自助采样有关，不太懂 ''' function to evaluate a CAD algorithm @param seriesUIDs: list of the seriesUIDs of the cases to be processed @param results_filename: file with results @param outputDir: output directory @param allNodules: dictionary with all nodule annotations of all cases, keys of the dictionary are the seriesuids @param CADSystemName: name of the CAD system, to be used in filenames and on FROC curve ''' nodOutputfile = open(os.path.join(outputDir,'CADAnalysis.txt'),'w') #写入CADAnalysis文件 nodOutputfile.write("\n") nodOutputfile.write((60 * "*") + "\n") nodOutputfile.write("CAD Analysis: %s\n" % CADSystemName) nodOutputfile.write((60 * "*") + "\n") nodOutputfile.write("\n") results = csvTools.readCSV(results_filename)#读取检测结果csv格式，seriesuid,coordX,coordY,coordZ,probability allCandsCAD = {} for seriesuid in seriesUIDs: #对每个病例读取相应的候选结节 # collect candidates from result file nodules = {} #从检测结果文件中获取与病例对应的候选结节，添加进字典 header = results[0] i = 0 for result in results[1:]: nodule_seriesuid = result[header.index(seriesuid_label)] if seriesuid == nodule_seriesuid: nodule = getNodule(result, header) nodule.candidateID = i nodules[nodule.candidateID] = nodule i += 1 if (maxNumberOfCADMarks > 0): # number of CAD marks, only keep must suspicous marks if len(nodules.keys()) > maxNumberOfCADMarks: # make a list of all probabilities probs = [] for keytemp, noduletemp in nodules.iteritems(): probs.append(float(noduletemp.CADprobability)) probs.sort(reverse=True) # sort from large to small probThreshold = probs[maxNumberOfCADMarks] nodules2 = {} nrNodules2 = 0 for keytemp, noduletemp in nodules.iteritems(): if nrNodules2 >= maxNumberOfCADMarks: break if float(noduletemp.CADprobability) > probThreshold: nodules2[keytemp] = noduletemp nrNodules2 += 1 nodules = nodules2 # print 'adding candidates: ' + seriesuid allCandsCAD[seriesuid] = nodules #将病例与对应候选结节存入字典 # open output files nodNoCandFile = open(os.path.join(outputDir, "nodulesWithoutCandidate_%s.txt" % CADSystemName), 'w') # --- iterate over all cases (seriesUIDs) and determine how # often a nodule annotation is not covered by a candidate # initialize some variables to be used in the loop candTPs = 0 #这里是一堆定义，让人头大 candFPs = 0 candFNs = 0 candTNs = 0 totalNumberOfCands = 0 #总候选结节数，也就是你的检测结果文件中的所有结节数量 totalNumberOfNodules = 0 #总标签结节数 doubleCandidatesIgnored = 0 irrelevantCandidates = 0 minProbValue = -1000000000.0 # minimum value of a float FROCGTList = [] FROCProbList = [] FPDivisorList = [] excludeList = [] FROCtoNoduleMap = [] ignoredCADMarksList = [] # -- loop over the cases for seriesuid in seriesUIDs: # get the candidates for this case try: candidates = allCandsCAD[seriesuid] except KeyError: candidates = {} # add to the total number of candidates totalNumberOfCands += len(candidates.keys()) # make a copy in which items will be deleted candidates2 = candidates.copy() #对候选结节复制一个副本 # get the nodule annotations on this case try: noduleAnnots = allNodules[seriesuid] #获取所有结节中属于该病例的结节（既包括真的结节，也包括无关的结节） except KeyError: noduleAnnots = [] # - loop over the nodule annotations for noduleAnnot in noduleAnnots: #对标签结节循环处理 # increment the number of nodules if noduleAnnot.state == "Included": #如果是用来评测的真结节，则 totalNumberOfNodules += 1 x = float(noduleAnnot.coordX) #获取结节坐标 y = float(noduleAnnot.coordY) z = float(noduleAnnot.coordZ) # 2. Check if the nodule annotation is covered by a candidate # A nodule is marked as detected when the center of mass of the candidate is within a distance R of # the center of the nodule. In order to ensure that the CAD mark is displayed within the nodule on the # CT scan, we set R to be the radius of the nodule size. diameter = float(noduleAnnot.diameter_mm) if diameter < 0.0: diameter = 10.0 radiusSquared = pow((diameter / 2.0), 2.0) found = False noduleMatches = [] for key, candidate in candidates.iteritems(): #对每一个候选结节，判断是否与真实结节相交 x2 = float(candidate.coordX) y2 = float(candidate.coordY) z2 = float(candidate.coordZ) dist = math.pow(x - x2, 2.) + math.pow(y - y2, 2.) + math.pow(z - z2, 2.) if dist < radiusSquared: #判断是否在半径距离内 if (noduleAnnot.state == "Included"): #若是用来检测的结节，添加进noduleMatches，并删除该候选结节 found = True noduleMatches.append(candidate) if key not in candidates2.keys():#这里不复杂，就是说把每个与标签结节相交的候选结节提取出来后，要删除副本中的id，这样是检测会否有其它标签结节也与该结节所属的候选结节相交 print "This is strange: CAD mark %s detected two nodules! Check for overlapping nodule annotations, SeriesUID: %s, nodule Annot ID: %s" % (str(candidate.id), seriesuid, str(noduleAnnot.id)) else: del candidates2[key] elif (noduleAnnot.state == "Excluded"): # an excluded nodule #若是无关结节，则添加进ignoredCADMarksList if bOtherNodulesAsIrrelevant: # delete marks on excluded nodules so they don't count as false positives if key in candidates2.keys(): irrelevantCandidates += 1 ignoredCADMarksList.append("%s,%s,%s,%s,%s,%s,%.9f" % (seriesuid, -1, candidate.coordX, candidate.coordY, candidate.coordZ, str(candidate.id), float(candidate.CADprobability))) del candidates2[key] if len(noduleMatches) > 1: # double detection #如果一个标签结节对应多个候选结节，记录下数目 doubleCandidatesIgnored += (len(noduleMatches) - 1) if noduleAnnot.state == "Included": #若该结节是include结节 # only include it for FROC analysis if it is included # otherwise, the candidate will not be counted as FP, but ignored in the # analysis since it has been deleted from the nodules2 vector of candidates if found == True: #若找到与之匹配的候选结节 # append the sample with the highest probability for the FROC analysis maxProb = None for idx in range(len(noduleMatches)): candidate = noduleMatches[idx] if (maxProb is None) or (float(candidate.CADprobability) > maxProb): maxProb = float(candidate.CADprobability) #记录匹配的候选结节的概率，将最大概率存入maxProb FROCGTList.append(1.0) #添加 1 FROCProbList.append(float(maxProb)) #添加刚刚的最大概率 FPDivisorList.append(seriesuid) #添加病例id excludeList.append(False) #添加False FROCtoNoduleMap.append("%s,%s,%s,%s,%s,%.9f,%s,%.9f" % (seriesuid, noduleAnnot.id, noduleAnnot.coordX, noduleAnnot.coordY, noduleAnnot.coordZ, float(noduleAnnot.diameter_mm), str(candidate.id), float(candidate.CADprobability))) candTPs += 1 #添加1 else: #若未找到与之匹配的结节 candFNs += 1 # append a positive sample with the lowest probability, such that this is added in the FROC analysis FROCGTList.append(1.0) FROCProbList.append(minProbValue) FPDivisorList.append(seriesuid) excludeList.append(True) FROCtoNoduleMap.append("%s,%s,%s,%s,%s,%.9f,%s,%s" % (seriesuid, noduleAnnot.id, noduleAnnot.coordX, noduleAnnot.coordY, noduleAnnot.coordZ, float(noduleAnnot.diameter_mm), int(-1), "NA")) nodNoCandFile.write("%s,%s,%s,%s,%s,%.9f,%s\n" % (seriesuid, noduleAnnot.id, noduleAnnot.coordX, noduleAnnot.coordY, noduleAnnot.coordZ, float(noduleAnnot.diameter_mm), str(-1))) # add all false positives to the vectors for key, candidate3 in candidates2.iteritems(): #此时candidata2中都是无人领取的结节，都是FP candFPs += 1 FROCGTList.append(0.0) FROCProbList.append(float(candidate3.CADprobability)) FPDivisorList.append(seriesuid) excludeList.append(False) FROCtoNoduleMap.append("%s,%s,%s,%s,%s,%s,%.9f" % (seriesuid, -1, candidate3.coordX, candidate3.coordY, candidate3.coordZ, str(candidate3.id), float(candidate3.CADprobability))) if not (len(FROCGTList) == len(FROCProbList) and len(FROCGTList) == len(FPDivisorList) and len(FROCGTList) == len(FROCtoNoduleMap) and len(FROCGTList) == len(excludeList)): nodOutputfile.write("Length of FROC vectors not the same, this should never happen! Aborting..\n") nodOutputfile.write("Candidate detection results:\n") nodOutputfile.write(" True positives: %d\n" % candTPs) nodOutputfile.write(" False positives: %d\n" % candFPs) nodOutputfile.write(" False negatives: %d\n" % candFNs) nodOutputfile.write(" True negatives: %d\n" % candTNs) nodOutputfile.write(" Total number of candidates: %d\n" % totalNumberOfCands) nodOutputfile.write(" Total number of nodules: %d\n" % totalNumberOfNodules) nodOutputfile.write(" Ignored candidates on excluded nodules: %d\n" % irrelevantCandidates) nodOutputfile.write(" Ignored candidates which were double detections on a nodule: %d\n" % doubleCandidatesIgnored) if int(totalNumberOfNodules) == 0: nodOutputfile.write(" Sensitivity: 0.0\n") else: nodOutputfile.write(" Sensitivity: %.9f\n" % (float(candTPs) / float(totalNumberOfNodules))) nodOutputfile.write(" Average number of candidates per scan: %.9f\n" % (float(totalNumberOfCands) / float(len(seriesUIDs)))) # compute FROC print FROCGTList print FROCProbList print len(FROCGTList) fps, sens, thresholds = computeFROC(FROCGTList,FROCProbList,len(seriesUIDs),excludeList) #这一步最为关键，计算FROC值，其实返回的是召回率与假阳性率的列表 if performBootstrapping: #是否使用自助采样，不太懂 fps_bs_itp,sens_bs_mean,sens_bs_lb,sens_bs_up = computeFROC_bootstrap(FROCGTList,FROCProbList,FPDivisorList,seriesUIDs,excludeList, numberOfBootstrapSamples=numberOfBootstrapSamples, confidence = confidence) # Write FROC curve #计算到上面就结束了，后面是一些画图 with open(os.path.join(outputDir, "froc_%s.txt" % CADSystemName), 'w') as f: for i in range(len(sens)): f.write("%.9f,%.9f,%.9f\n" % (fps[i], sens[i], thresholds[i])) # Write FROC vectors to disk as well with open(os.path.join(outputDir, "froc_gt_prob_vectors_%s.csv" % CADSystemName), 'w') as f: for i in range(len(FROCGTList)): f.write("%d,%.9f\n" % (FROCGTList[i], FROCProbList[i])) fps_itp = np.linspace(FROC_minX, FROC_maxX, num=10001) sens_itp = np.interp(fps_itp, fps, sens) frvvlu = 0 nxth = 0.125 for fp, ss in zip(fps_itp, sens_itp): if abs(fp - nxth) < 3e-4: frvvlu += ss nxth *= 2 if abs(nxth - 16) < 1e-5: break print(frvvlu/7, nxth) print(sens_itp[fps_itp==0.125]+sens_itp[fps_itp==0.25]+sens_itp[fps_itp==0.5]+sens_itp[fps_itp==1]+sens_itp[fps_itp==2]\ +sens_itp[fps_itp==4]+sens_itp[fps_itp==8]) if performBootstrapping: # Write mean, lower, and upper bound curves to disk with open(os.path.join(outputDir, "froc_%s_bootstrapping.csv" % CADSystemName), 'w') as f: f.write("FPrate,Sensivity[Mean],Sensivity[Lower bound],Sensivity[Upper bound]\n") for i in range(len(fps_bs_itp)): f.write("%.9f,%.9f,%.9f,%.9f\n" % (fps_bs_itp[i], sens_bs_mean[i], sens_bs_lb[i], sens_bs_up[i])) else: fps_bs_itp = None sens_bs_mean = None sens_bs_lb = None sens_bs_up = None # create FROC graphs if int(totalNumberOfNodules) > 0: graphTitle = str("") fig1 = plt.figure() ax = plt.gca() clr = 'b' plt.plot(fps_itp, sens_itp, color=clr, label="%s" % CADSystemName, lw=2) if performBootstrapping: plt.plot(fps_bs_itp, sens_bs_mean, color=clr, ls='--') plt.plot(fps_bs_itp, sens_bs_lb, color=clr, ls=':') # , label = "lb") plt.plot(fps_bs_itp, sens_bs_up, color=clr, ls=':') # , label = "ub") ax.fill_between(fps_bs_itp, sens_bs_lb, sens_bs_up, facecolor=clr, alpha=0.05) xmin = FROC_minX xmax = FROC_maxX plt.xlim(xmin, xmax) plt.ylim(0.5, 1) plt.xlabel('Average number of false positives per scan') plt.ylabel('Sensitivity') plt.legend(loc='lower right') plt.title('FROC performance - %s' % (CADSystemName)) if bLogPlot: plt.xscale('log', basex=2) ax.xaxis.set_major_formatter(FixedFormatter([0.125,0.25,0.5,1,2,4,8])) # set your ticks manually ax.xaxis.set_ticks([0.125,0.25,0.5,1,2,4,8]) ax.yaxis.set_ticks(np.arange(0.5, 1, 0.1)) # ax.yaxis.set_ticks(np.arange(0, 1.1, 0.1)) plt.grid(b=True, which='both') plt.tight_layout() plt.savefig(os.path.join(outputDir, "froc_%s.png" % CADSystemName), bbox_inches=0, dpi=300) return (fps, sens, thresholds, fps_bs_itp, sens_bs_mean, sens_bs_lb, sens_bs_up)
这里会计算FROC，调用的是computeFROC函数。
def computeFROC(FROCGTList, FROCProbList, totalNumberOfImages, excludeList): # Remove excluded candidates FROCGTList_local = [] FROCProbList_local = [] for i in range(len(excludeList)): if excludeList[i] == False: FROCGTList_local.append(FROCGTList[i]) FROCProbList_local.append(FROCProbList[i]) numberOfDetectedLesions = sum(FROCGTList_local) totalNumberOfLesions = sum(FROCGTList) totalNumberOfCandidates = len(FROCProbList_local) fpr, tpr, thresholds = skl_metrics.roc_curve(FROCGTList_local, FROCProbList_local, pos_label=1) if sum(FROCGTList) == len(FROCGTList): # Handle border case when there are no false positives and ROC analysis give nan values. print "WARNING, this system has no false positives.." fps = np.zeros(len(fpr)) else: fps = fpr * (totalNumberOfCandidates - numberOfDetectedLesions) / totalNumberOfImages sens = (tpr * numberOfDetectedLesions) / totalNumberOfLesions return fps, sens, thresholds
到此，结束。
