def TraMapSag(sagPoints,traPath):
r1, c1, o1, pixelSpacingXY, n1 = sagPoints
orienta, pos, pixelSpa = dicom_metainfo(traPath,
['0020|0037', '0020|0032', '0028|0030'])
(r2, c2), o2, pixelSpa = parse_orentation(orienta), parse_position(pos), parse_pixelSpacing(
pixelSpa)
image = dicom2array(traPath)
srcHeigh, srcWidth = image.shape[:2]
p1 = o2 # 左上角的点
p2 = p1 + r2 * pixelSpa[0] * (srcWidth - 1) # 右上角的点
p3 = p2 + c2 * pixelSpa[1] * (srcHeigh - 1) # 右下角的点
p4 = p1 + c2 * pixelSpa[1] * (srcHeigh - 1) # 左下角的点
# 点到平面的距离
# p1 O1
dv1 = np.dot(p1 - o1, n1)
# p2 O1
dv2 = np.dot(p2 - o1, n1)
# p3 O1
dv3 = np.dot(p3 - o1, n1)
# p4 O1
dv4 = np.dot(p4 - o1, n1)
iscross12 = ((dv1 > 0) & (dv2 < 0)) | ((dv1 < 0) & (dv2 > 0))
iscross23 = ((dv2 > 0) & (dv3 < 0)) | ((dv2 < 0) & (dv3 > 0))
iscross34 = ((dv3 > 0) & (dv4 < 0)) | ((dv3 < 0) & (dv4 > 0))
iscross41 = ((dv4 > 0) & (dv1 < 0)) | ((dv4 < 0) & (dv1 > 0))
if not (iscross12 | iscross23 | iscross34 | iscross41):
return
cps = []
for iscross, (pv, pp), (ddv, ddp) in zip([iscross12, iscross23, iscross34, iscross41],
[(p1, p2), (p2, p3), (p3, p4), (p4, p1)],
[(dv1, dv2), (dv2, dv3), (dv3, dv4), (dv4, dv1)]):
if not iscross:
# 是否相交
continue
# 计算交点坐标 相似三角形？
cp = [pv[i] + (pp[i] - pv[i]) * ddv / (ddv - ddp) for i in range(3)]
cps.append(np.array(cp))
# for cp in cps:
# # 看下算出来的点在不在定位图平面上 带入平面方程 正确应该是几乎等于0
# print(n1 * (cp - o1))
assert len(cps) == 2
coords = []
coords_img = []
for cp in cps:
# 投影
xmap = np.dot(cp - o1, r1) / np.linalg.norm(r1)
ymap = np.dot(cp - o1, c1) / np.linalg.norm(c1)
# 这个没什么用
coords.append(np.array(xmap, ymap))
# pixel坐标
coords_img.append(np.array([int(xmap / pixelSpacingXY[0]), int(ymap / pixelSpacingXY[1])])) #得到映射到sag的两个点
return coords_img
def SagKeyPoints(sagPath):
image1 = dicom2array(sagPath)
orienta, pos, pixelSpacingXY = dicom_metainfo(sagPath,
['0020|0037', '0020|0032', '0028|0030'])
(r1, c1), o1, pixelSpacingXY = parse_orentation(orienta), parse_position(pos), parse_pixelSpacing(
pixelSpacingXY)
srcHeigh, srcWidth = image1.shape[:2]
n1 = np.cross(r1, c1)
p1 = o1 # 左上角的点
p2 = p1 + r1 * pixelSpacingXY[0] * (srcWidth - 1) # 右上角的点
p3 = p2 + c1 * pixelSpacingXY[1] * (srcHeigh - 1) # 右下角的点
p4 = p1 + c1 * pixelSpacingXY[1] * (srcHeigh - 1) # 左下角的点
return r1,c1,o1,pixelSpacingXY,n1
def to_Json(pointsPath, dataTestPath, Tojson=r"reslut_.json"):
jsonFile = []
identificationList = [
"L5-S1", "L5", "L4-L5", "L4", "L3-L4", "L3", "L2-L3", "L2", "L1-L2",
"L1", "T12-L1", "T12"
]
f = open(pointsPath, 'r')
datas = f.readlines()
count = 0
for data in datas:
Temp = {
"studyUid":
"1.2.840.473.8013.20181026.1091511.864.22434.92", # dicom study UID
"data": [{
"instanceUid":
"1.3.46.670589.11.32898.5.0.16316.2018102610554807009", # dicom instance UID
"seriesUid":
"1.3.46.670589.11.32898.5.0.16316.2018102610554554000", # dicom series UID
"annotation": [{
"data": [{
"annotator":
0, # 可选
"point": [ # 关键点标注
{
"coord": [252, 435], # 点的像素坐标
"tag": {
"disc": "v2", # 椎间盘类型膨出
"identification": "L1-L2" # 椎间盘定位L1-L2间椎间盘
},
"zIndex": 0, # 第几个slice
},
]
}],
}]
}]
}
jsonTemp = Temp
d = data.split()
ImgPath = d[0].split('/')[-1].split('.')[0].split('_') #寻找dcm
dicmPath = dataTestPath + '/' + ImgPath[0] + "/" + ImgPath[1] + '.dcm'
studyUid, seriesUid, instanceUid, instanceNumber = readyData.dicom_metainfo(
dicmPath, ['0020|000d', '0020|000e', '0008|0018', '0020|0013'])
jsonTemp["studyUid"] = studyUid
jsonTemp["data"][0]["instanceUid"] = instanceUid
jsonTemp["data"][0]["seriesUid"] = seriesUid
jsonTemp["data"][0]["annotation"] = []
annotationDic = {}
annotationDic["data"] = {}
annotationDic["data"]["point"] = []
if len(d) > 1:
d = sortPoint(d)
for index, target in enumerate(d):
tagDic = {}
point_dic = {}
x, y, vd, c, p = target.split(',')
# c = "v2"
point_dic["coord"] = [int(x), int(y)]
if vd == "disc":
if p == '1':
tagDic[vd] = c
else:
tagDic[vd] = c
else:
tagDic[vd] = c
# tagDic["p"] = p
tagDic["identification"] = identificationList[index]
point_dic["tag"] = tagDic
point_dic["zIndex"] = instanceNumber
annotationDic["data"]["point"].append(point_dic)
if index == 11:
break
jsonTemp["data"][0]["annotation"].append(annotationDic)
jsonFile.append(jsonTemp)
count += 1
f.close()
f1 = open(Tojson, "w")
json.dump(jsonFile, f1)
f1.close()
print("json中数量:", count)
def metricDetect(yolo, truejson, testFile, sliceResize):
with open(truejson, "r", encoding="utf-8") as f:
truejsonTarge = json.loads(f.read())
from classifycation import CNN
cnn_disc = CNN(sliceResize)
model_disc = cnn_disc.loadTestdisc()
cnn_discv5 = CNN(sliceResize)
model_discv5 = cnn_discv5.loadTestdiscv5()
cnn_vertebra = CNN(sliceResize)
model_vertebra = cnn_vertebra.loadTestvertebra()
TruePointNum = 0
TP = 0 #在目标内且 如果分类正确
FP = 0 #在目标内且 分类错误 and 如果预测的点不落在任何标注点
FN = 0 #标注点没有被任何预测点正确命中
vertebraTP = 0
vertebraFP = 0
discTP = 0
discFP = 0
mm6c = 0
mmc = 0
count = 0
study = os.listdir(testFile)
for studyI in study:
Impath = os.listdir(os.path.join(testFile, studyI))
for Impathi in Impath:
try:
studyUid, seriesUid, instanceUid = dicom_metainfo(
os.path.join(os.path.join(testFile, studyI), Impathi),
['0020|000d', '0020|000e', '0008|0018'])
for studyid in truejsonTarge:
if studyUid == studyid["studyUid"] and seriesUid == studyid[
"data"][0]["seriesUid"] and instanceUid == studyid[
"data"][0]["instanceUid"]:
count += 1
image = dicom2array(
os.path.join(os.path.join(testFile, studyI),
Impathi))
truepoints = studyid["data"][0]["annotation"][0][
"data"]["point"]
image = Image.fromarray(image)
image = image.convert("RGB")
r_image, points = yolo.detect_image(image)
#判断检测点与真实点
mean = np.mean(image)
std = np.mean(np.square(image - mean))
image = (image - mean) / np.sqrt(std)
image = np.asarray(image)
m, n, _ = image.shape
matchedIndex = []
for truepoint in truepoints:
mmc += 1
xt, yt = list(map(int, truepoint["coord"]))
tag = truepoint["tag"]
if "disc" in tag.keys():
ct = tag["disc"]
else:
ct = tag["vertebra"]
xy = sorted(points, key=lambda x: x[2])
deta = []
for i in range(len(xy) - 1):
deta.append(xy[i + 1][2] - xy[i][2])
deta = np.mean(deta)
print("deta:", deta)
for index, metricpoint in enumerate(points):
c, x, y = metricpoint
mm = 6
mm8 = 8
if x <= xt + mm and x >= xt - mm and y <= yt + mm and y >= yt - mm: # 在标注点内
matchedIndex.append(index)
mm6c += 1
TruePointNum += 1
w1, h1 = m / 5, n / 20 # 椎间盘要求更长而不是更高 n控制高度
offset = 5
miny = y - deta
maxy = y + deta
minx = x + offset - w1 / 2
maxx = x + offset + w1 / 2
# nimg_x = image[int(miny):int(maxy), int(minx):int(maxx)]
if "disc" in tag.keys():
xlen2 = int(w1 / 2)
X = dealImg(image, int(miny),
int(maxy),
int(minx) + int(xlen2),
int(maxx), sliceResize)
X = X.reshape([1] + sliceResize + [1])
p = model_disc.predict(X)
pre = np.argmax(p, axis=1)
pc = cnn_disc.disc_label[int(pre)]
if pc == ct:
TP += 1
discTP += 1
else:
FP += 1
discFP += 1
else:
w1, h1 = m / 5, n / 20 # 识别框的宽度和高度 更大
offset = 0
miny = y - h1 #- deta - offset
maxy = y + h1 #+ deta + offset
minx = x + offset - w1 / 2
maxx = x + offset + w1 / 2
# nimg_x = image[int(miny):int(maxy), int(minx):int(maxx)]
# nimg_x = cv2.resize(nimg_x, (sliceResize[1], sliceResize[0]))[:, :, 1]
X = dealImg(image, int(miny),
int(maxy), int(minx),
int(maxx), sliceResize)
# plt.subplot(4, 5, int(1 + 5 * 1))
# plt.imshow(X[:,:,0])
# plt.show()
X = X.reshape([1] + sliceResize + [1])
p = model_vertebra.predict(X)
pre = np.argmax(p, axis=1)
pc = cnn_vertebra.vertebra_label[int(
pre)]
if pc == ct:
TP += 1
vertebraTP += 1
else:
FP += 1
vertebraFP += 1
break
FP += len(points) - len(matchedIndex)
except:
print("文件错误：",
os.path.join(os.path.join(dataPath, studyI), Impathi))
print("总共点数:", mmc)
print("TP/总定位:", TruePointNum / mmc, "TP/(TP+FP):", TP / (TP + FP + 1e-2))
print("discTP/(TP+FP):", discTP / (discTP + discFP), "vertebraTP/(TP+FP):",
vertebraTP / (vertebraTP + vertebraFP))
def step1(dataPath, jsondata, truejsonPath=None):
'''
将 图片 与 锚点对应
:return:
'''
image = {}
count = 0
study = os.listdir(dataPath)
for studyI in study:
Impath = os.listdir(os.path.join(dataPath, studyI))
# print("study:", studyI)
for Impathi in Impath:
try:
studyUid, seriesUid, instanceUid, desc, idex = dicom_metainfo(
os.path.join(os.path.join(dataPath, studyI), Impathi), [
'0020|000d', '0020|000e', '0008|0018', '0008|103e',
"0020|0013"
])
for studyid in jsondata:
if studyUid == studyid["studyUid"] and seriesUid == studyid[
"data"][0]["seriesUid"] and instanceUid == studyid[
"data"][0]["instanceUid"]:
count += 1
image[studyI + "_" + Impathi] = dicom2array(
os.path.join(os.path.join(dataPath, studyI),
Impathi))
# 在yolo.py中运行watchTest打开注释
img = dicom2array(
os.path.join(os.path.join(dataPath, studyI),
Impathi))
# if studyI == seestudy:
plt.figure()
plt.title(studyI + "_" + Impathi + "\t" + desc + '\t' +
idex)
plt.imshow(img)
if truejsonPath != None:
for studyid1 in truejsonPath:
if studyUid == studyid1[
"studyUid"] and seriesUid == studyid1[
"data"][0][
"seriesUid"] and instanceUid == studyid1[
"data"][0]["instanceUid"]:
truePoints = studyid1["data"][0][
"annotation"][0]["data"]["point"]
for point in truePoints:
x, y = point['coord']
print("True zIndex:", point["zIndex"])
tag = point['tag']
loc = tag['identification']
if "disc" in tag.keys():
k = tag["disc"]
else:
k = tag["vertebra"]
# if studyI == seestudy:
plt.scatter(x, y, s=12, c='g')
plt.text(x=x + 5,
y=y - 5,
s=loc + '**' + k)
for point in studyid["data"][0]["annotation"][0][
"data"]["point"]:
x, y = point['coord']
# print("zIndex:",point["zIndex"] ," red")
tag = point['tag']
loc = tag['identification']
if "disc" in tag.keys():
k = tag["disc"]
else:
k = tag["vertebra"]
# if studyI == seestudy:
plt.scatter(x, y, s=6, c='r')
plt.text(x=x + 5, y=y, s=loc + '--' + k)
# if studyI == seestudy:
plt.show()
except:
print("文件错误：",
os.path.join(os.path.join(dataPath, studyI), Impathi))
def step1Test(dataPath,testTxt=None):
'''
在准备jpg时，首先先判断是否为T2_sag，只对T2_sag做定位识别
target: t2sag:0, t1sag:1, t2tra:2
:param dataPath:
:return:
'''
count = 0
study = os.listdir(dataPath)
f = open(testTxt, 'r')
datas = f.readlines()
f.close()
for data in datas:
d = data.split()
ImgPath = d[0].split('/')[-1].split('.')[0].split('_') # 寻找dcm
studyI = ImgPath[0]
dicmPath = dataPath + '/' + ImgPath[0] + "/" + ImgPath[1] + '.dcm' #answer中间帧
Impath = os.listdir(os.path.join(dataPath, studyI)) #找到study下的所有dcm 为了寻找tra，同一标签下第一左边方差小
describes2pos = {}
describes2imgPathes = {}
for Impathi in Impath:
pos = dicom_metainfo(os.path.join(os.path.join(dataPath, studyI), Impathi), ['0020|0032'])[0]
describ = dicom_metainfo(os.path.join(os.path.join(dataPath, studyI), Impathi), ['0008|103e'])[0]
seriesNumber = dicom_metainfo(os.path.join(os.path.join(dataPath, studyI), Impathi), ['0020|0011'])[0]
describ += seriesNumber
if re.match(r"(.*)[sS][cC][oO][uU][tT](.*)", describ, 0): #不加入
continue
elif re.match(r"(.*)[sS][aA][gG](.*)",describ,flags=0) or re.match(r"(.*)[sS][aA][gG](.*)",describ,0):
continue
else:
if describ not in describes2pos.keys():
describes2pos[describ] = [parse_position(pos)[0]]
describes2imgPathes[describ] = [Impathi]
else:
describes2pos[describ].append(parse_position(pos)[0])
describes2imgPathes[describ].append(Impathi)
describes2posN = {}
describes2imgPathesN = {}
flagN = False
for key, value in describes2pos.items():
if re.match(r"(.*)[tT][rR][aA](.*)",key,flags=0):
flagN = True
describes2posN[key] = value
describes2imgPathesN[key] = describes2imgPathes[key]
if flagN:
describes2pos = describes2posN
describes2imgPathes = describes2imgPathesN
plt.figure(1)
imgSAG = dicom2array(dicmPath)
plt.imshow(imgSAG)
sagPoints = SagKeyPoints(dicmPath)
index = 0
for key,value in describes2pos.items():
if np.std(np.array(value))<10 and len(value) > 6:
# plt.figure(2)
print("找到tra! -- ",studyI,key," -- ",len(value))
count += 1
clusData = []
for traPath in describes2imgPathes[key]:
traDim = os.path.join(os.path.join(dataPath, studyI), traPath)
TraMapSagPoints = TraMapSag(sagPoints,traDim)
imgTRA = dicom2array(traDim)
# plt.subplot(2,int(len(value)/2)+1,index+1)
# plt.imshow(imgTRA)
plt.plot((TraMapSagPoints[0][0], TraMapSagPoints[1][0]), (TraMapSagPoints[0][1], TraMapSagPoints[1][1]))
x1,x2 = int(TraMapSagPoints[0][0]),int(TraMapSagPoints[1][0])
y1,y2 = int(TraMapSagPoints[0][1]),int(TraMapSagPoints[1][1])
clusData.append([y1,y2,x1,x2, traDim])
plt.plot(x1, y1, "o", color='red', ms=10)
plt.plot(x2, y2, "o", color='g', ms=10)
plt.plot(int((x1+x2)/2), int((y1+y2)/2), "o", color='g', ms=10)
plt.text(x1, y1,str(index))
index += 1
thresh = 30
print(sorted(clusData,key=lambda x:x[0]))
clusData = sorted(clusData, key=lambda x: x[0])
clusDataN = [[i[0]] for i in clusData]
traDimN = [i[-1] for i in clusData]
clusters = hcluster.fclusterdata(np.array(clusDataN), thresh, criterion="distance")
print(clusters,set(clusters.tolist()))
clustersDic = dealCluster(clusters,clusData)
print(clustersDic)
d = sortPoint(d)
for index, target in enumerate(d):
x, y, vd, c, p = target.split(',')
if vd == "disc":
print("==")
flag,dicm = xyIndcm(int(x),int(y),clustersDic)
if flag:
print(dicm)
plt.plot(int(x), int(y), "o", color='r', ms=5)
plt.show()
print(count)