def get_grad_img(polar_patch):
OFFX = 2
polar_patch_gaussian = gaussian_filter(polar_patch, sigma=5)
sy = ndimage.sobel(polar_patch_gaussian, axis=0, mode='constant')
sx = ndimage.sobel(polar_patch_gaussian, axis=1, mode='constant')
polar_grad = np.hypot(sx, sy)
polar_grad = croppatch(croppatch(polar_grad, 128, 128, 127, 127), 127,
127 - OFFX, 128, 128)
polar_grad[0] = polar_grad[1]
polar_grad[-1] = polar_grad[-2]
gradimg = polar_grad #/ np.max(polar_grad)
return gradimg
def multi_min_dist_map(min_dist_cnn, dicomslice, octy, octx, patchheight,
patchwidth):
search_patch_rz_batch = []
stride = 10
rg = 2
DEBUG = 0
for ofi in np.arange(-rg, rg + 1):
for ofj in np.arange(-rg, rg + 1):
octyof = octy + ofi * stride
octxof = octx + ofj * stride
searchpatch = croppatch(dicomslice, octyof, octxof, patchheight,
patchwidth)
searchpatchrz = cv2.resize(searchpatch, (0, 0), fx=4, fy=4)
searchpatchrz = searchpatchrz / np.max(searchpatchrz)
# searchpatchrz = searchpatch
search_patch_rz_batch.append(searchpatchrz)
'''plt.imshow(searchpatchrz)
plt.title('Original Patch'+str(octyof)+str(octxof))
plt.show()'''
search_patch_rz_batch = np.array(search_patch_rz_batch)
min_dist_search_patch_batch = min_dist_cnn.predict(
search_patch_rz_batch[:, :, :, None])[:, :, :, 0]
min_dist_search_patch = []
for i in range(min_dist_search_patch_batch.shape[0]):
ofi = i // (2 * rg + 1) - rg
ofj = i % (2 * rg + 1) - rg
# print(ofi,ofj)
offpred = croppatch(min_dist_search_patch_batch[i],
256 - ofi * stride * 4, 256 - ofj * stride * 4,
256, 256)
# offpred = croppatch(min_dist_search_patch_batch[i], 256 - ofi * stride, 256 - ofj * stride, 256, 256)
if DEBUG:
plt.suptitle('Original Patch' + str(ofi) + str(ofj))
plt.subplot(1, 3, 1)
plt.imshow(search_patch_rz_batch[i])
plt.subplot(1, 3, 2)
plt.imshow(min_dist_search_patch_batch[i])
plt.subplot(1, 3, 3)
plt.imshow(offpred)
#plt.colorbar()
plt.show()
print(np.max(offpred))
min_dist_search_patch.append(offpred / np.max(offpred))
min_dist_search_patch = np.array(min_dist_search_patch)
min_dist_search_patch = np.max(min_dist_search_patch, axis=0)
# plt.imshow(min_dist_search_patch)
return min_dist_search_patch
def multi_min_dist_pred_component(min_dist_cnn,
dicomslice,
bb,
patchheight,
patchwidth,
kernelmask=None):
#min_dist_search_patch = multi_min_dist_map(min_dist_cnn, dicomslice, octy, octx, patchheight, patchwidth)
cts, nms_dist_map_p = multi_min_dist_pred_withinbb(min_dist_cnn,
dicomslice, bb, 64, 64)
min_dist_search_patch = nms_dist_map_p[:, :, 0]
if kernelmask is None:
kernelmask = gen_2d_gaussion(min_dist_search_patch.shape[0], 0.5)
component_bbs, nms_dist_map_p = find_con_region(min_dist_search_patch *
kernelmask,
fig=1)
#prepare output img
search_patch_center = croppatch(dicomslice, bb.y, bb.x, patchheight,
patchwidth)
search_patch_center_rz = cv2.resize(search_patch_center, (0, 0),
fx=4,
fy=4)
# search_patch_center_rz = search_patch_center
# print('map max',np.max(search_patch_center_rz))
nms_dist_map_p[:, :, 2] = search_patch_center_rz * 255
return component_bbs, nms_dist_map_p
def multi_min_dist_pred(min_dist_cnn,
dicomslice,
octy,
octx,
patchheight,
patchwidth,
kernelmask=None):
min_dist_search_patch = multi_min_dist_map(min_dist_cnn, dicomslice, octy,
octx, patchheight, patchwidth)
if kernelmask is None:
kernelmask = gen_2d_gaussion(min_dist_search_patch.shape[0], 0.5)
dist_pred_ct, nms_dist_map_p = find_nms_center(min_dist_search_patch *
kernelmask,
fig=1)
search_patch_center = croppatch(dicomslice, octy, octx, patchheight,
patchwidth)
search_patch_center_rz = cv2.resize(search_patch_center, (0, 0),
fx=4,
fy=4)
# search_patch_center_rz = search_patch_center
#print('map max',np.max(search_patch_center_rz))
nms_dist_map_p[:, :, 2] = search_patch_center_rz * 255
cts = [[int(i.split('-')[0]), int(i.split('-')[1])]
for i in list(dist_pred_ct.keys())]
if len(cts) == 0:
print('no ct')
elif len(cts) != 1:
pass
#print('cts', len(cts))
return cts, nms_dist_map_p
def load_aug_patch(self, slicei, type):
if not self.valid_slicei(slicei):
return None
if type == 'polar_patch':
aug_filename = self.caselist['slices'][slicei]['augpatch'][
'polarpatchbatchname']
elif type == 'cart_patch':
aug_filename = self.caselist['slices'][slicei]['augpatch'][
'cartpatchbatchname']
elif type == 'polar_cont':
aug_filename = self.caselist['slices'][slicei]['augpatch'][
'polarcoutbatchname']
elif type == 'cart_label':
aug_offs = self.load_aug_off(slicei)
cart_label_vw = self.load_cart_vw(slicei)
cart_label_vw_aug_batch = []
for augi in range(len(aug_offs)):
ctx = aug_offs[augi][0]
cty = aug_offs[augi][1]
cart_label_vw_aug = croppatch(cart_label_vw, 256 + cty,
256 + ctx, 256, 256)
cart_label_vw_aug_batch.append(cart_label_vw_aug)
return cart_label_vw_aug_batch
elif type == 'polar_label':
aug_offs = self.load_aug_off(slicei)
cart_label_vw = self.load_cart_vw(slicei)
polar_label_vw_aug_batch = []
for augi in range(len(aug_offs)):
ctx = aug_offs[augi][0]
cty = aug_offs[augi][1]
cart_label_vw_aug = croppatch(cart_label_vw, 256 + cty,
256 + ctx, 256, 256)
polar_label_vw_aug = topolar(cart_label_vw_aug, 256, 256)
polar_label_vw_aug_batch.append(polar_label_vw_aug)
return polar_label_vw_aug_batch
else:
print('unknown type')
return
aug_filename = self.targetprefix + aug_filename
if not os.path.exists(aug_filename):
print('no aug polar patch file', aug_filename)
raise FileNotFoundError
aug_batch = np.load(aug_filename).astype(np.float)
return aug_batch
def intpath(pos1, pos2, dicomstack):
DEBUG = 0
pos1 = np.array(pos1)
pos2 = np.array(pos2)
pos1int = [int(round(posi)) for posi in pos1]
pos2int = [int(round(posi)) for posi in pos2]
direction = pos2 - pos1
dist = np.linalg.norm(direction)
dirnorm = direction / dist
intp = []
intp.append(dicomstack[pos1int[2]][pos1int[1]][pos1int[0]])
if DEBUG:
pdsp = croppatch(dicomstack[pos1int[2]], pos1int[1], pos1int[0])
pdsp[pdsp.shape[0] // 2] = np.max(pdsp)
pdsp[:, pdsp.shape[1] // 2] = np.max(pdsp)
plt.title('start')
plt.imshow(pdsp)
plt.show()
for stepi in range(1, int(np.floor(dist))):
cpos = pos1 + dirnorm * stepi
cposint = np.array([int(np.round(cposi)) for cposi in cpos])
if DEBUG:
pdsp = croppatch(dicomstack[cposint[2]], cpos[1], cpos[0])
pdsp[pdsp.shape[0] // 2] = np.max(pdsp)
pdsp[:, pdsp.shape[1] // 2] = np.max(pdsp)
plt.imshow(pdsp)
plt.title(str(cposint[2]))
plt.show()
valceil = dicomstack[int(np.ceil(cpos[2]))][cposint[1]][cposint[0]]
valfloor = dicomstack[int(np.floor(cpos[2]))][cposint[1]][cposint[0]]
w1 = int(np.ceil(cpos[2])) - cpos[2]
w2 = cpos[2] - int(np.floor(cpos[2]))
intp.append(valceil * w2 / (w1 + w2) + valfloor * w1 / (w1 + w2))
intp.append(dicomstack[pos2int[2]][pos2int[1]][pos2int[0]])
if DEBUG:
pdsp = croppatch(dicomstack[pos2int[2]], pos2int[1], pos2int[0])
pdsp[pdsp.shape[0] // 2] = np.max(pdsp)
pdsp[:, pdsp.shape[1] // 2] = np.max(pdsp)
plt.title('end')
plt.imshow(pdsp)
plt.show()
return intp
def crop_dcm_stack(dicomstack, sid, cty, ctx, hps, RESCALE, depth=3):
cartstack = croppatch(dicomstack[sid], cty, ctx, hps, hps)
cartstackrz = cv2.resize(cartstack, (0, 0), fx=RESCALE, fy=RESCALE)
cartstackrz = np.repeat(cartstackrz[:, :, None], depth, axis=2)
nei = depth // 2
for ni in range(1, nei + 1):
imgslicep = sid - ni
if imgslicep >= 0:
cartstack = croppatch(dicomstack[imgslicep], cty, ctx, 64, 64)
cartstackrz[..., nei - ni] = cv2.resize(cartstack, (0, 0),
fx=RESCALE,
fy=RESCALE)
imgslicen = sid + ni
if imgslicen < dicomstack.shape[0]:
cartstack = croppatch(dicomstack[imgslicen], cty, ctx, 64, 64)
cartstackrz[..., nei + ni] = cv2.resize(cartstack, (0, 0),
fx=RESCALE,
fy=RESCALE)
cartstackrz = cartstackrz / np.max(cartstackrz)
return cartstackrz
def cont_ct_within_bb(min_dist_cnn, dicomstack, slicei, bb, seq, posstart):
DEBUG = 0
if slicei == seq[-1][0]:
cts, nms_dist_map_p = multi_min_dist_pred_withinbb(
min_dist_cnn, dicomstack[slicei], bb, 64, 64)
else:
cts, nms_dist_map_p = multi_min_dist_pred(min_dist_cnn,
dicomstack[slicei], bb.y,
bb.x, 64, 64)
# merge multiple ct in the same connected region
merge_cts = mergects(nms_dist_map_p[:, :, 0], cts)
# convert ct to dicom coordinate
cts_dcm = to_dcm_cord(merge_cts, bb)
if DEBUG:
plt.imshow(nms_dist_map_p)
plt.show()
if len(cts_dcm) == 1:
bbx = cts_dcm[0][0]
bby = cts_dcm[0][1]
elif len(cts_dcm) > 1:
# last slice choose lower artery
if slicei == seq[-1][0]:
mapint = [
nms_dist_map_p[int(round(merge_cts[ica][1])),
int(round(merge_cts[ica][0])), 0]
for ica in range(len(merge_cts))
]
print('mapint', mapint)
#ypos = [cts_dcm[ica][1] for ica in range(len(cts_dcm))]
#ica = np.argmax(ypos)
ica = np.argmax(mapint)
bbx = cts_dcm[ica][0]
bby = cts_dcm[ica][1]
# other slices find minimum int change
else:
#print('multiple', cts_dcm,bbprev)
ctm = ct_with_min_change(posstart, slicei, cts_dcm, dicomstack)
#print('ctm', ctm)
bbx = ctm[0]
bby = ctm[1]
if DEBUG:
plt.imshow(croppatch(dicomstack[slicei], bby, bbx))
plt.show()
else:
print('no cts')
return bbx, bby
def displayvw(self,figfilename=None):
plt.figure(figsize=(18, 5))
for slicei in range(len(self.seg['segct'])):
if self.seg['segct'][slicei] is None or self.seg['cartcont']['lumen'][slicei] is None:
continue
ct = np.mean(self.seg['segct'][slicei],axis=0)
dcmpatch = croppatch(self.plotvw(slicei), ct[1], ct[0], 64, 64)
maxval = np.max(dcmpatch)
dcmpatch[:,dcmpatch.shape[1]//2] = maxval
dcmpatch[dcmpatch.shape[1] // 2] = maxval
plt.subplot(2, int(np.ceil(len(self.dicomstack) / 2)), slicei + 1)
plt.imshow(dcmpatch)
plt.title(slicei)
if figfilename is not None:
plt.savefig(figfilename)
else:
plt.show()
plt.close()
def load_dcm_stack(dicompath,
norm=0,
seq='101',
refscale=1,
shift_x=0,
shift_y=0):
piname = dicompath.split('/')[-2]
einame = dicompath.split('/')[-1]
imgnamepattern = os.path.join(dicompath, einame + 'S' + seq + 'I*.dcm')
targetimgs = glob.glob(imgnamepattern)
if len(targetimgs) == 0:
print('No dcm', imgnamepattern)
return
slices = [int(i.split('I')[-1][:-4]) for i in targetimgs]
print(slices, np.min(slices), np.max(slices))
cartimgstack = np.zeros((np.max(slices), 512, 512))
imgnamepattern = os.path.join(dicompath, einame + 'S' + seq + 'I%d.dcm')
if np.min(slices) != 1:
print('min slice not 1')
for slicei in range(np.min(slices), np.max(slices) + 1):
imgfilename = imgnamepattern % slicei
if os.path.exists(imgfilename):
RefDs = pydicom.read_file(imgfilename).pixel_array
if RefDs.shape != (512, 512):
# print('Not regular size',RefDs.shape)
if RefDs.shape[0] == RefDs.shape[1]:
RefDs = cv2.resize(RefDs, (512, 512))
else:
padarr = np.zeros((max(RefDs.shape), max(RefDs.shape)))
padarr[0:RefDs.shape[0], 0:RefDs.shape[1]] = RefDs
RefDs = padarr
RefDs = cv2.resize(RefDs, (512, 512))
if refscale != 1:
RefDs = cv2.resize(RefDs, (0, 0), fx=refscale, fy=refscale)
if shift_y != 0 or shift_x != 0:
RefDs = croppatch(RefDs, (shift_y + 256), (shift_x + 256), 256,
256)
dcmimg = RefDs / np.max(RefDs)
cartimgstack[slicei - 1] = dcmimg
else:
print('no slice img for', imgfilename)
return cartimgstack
def display_tracklet_dcm_patch(tracklet, dicomstack, figfilename=None):
plt.figure(figsize=(18, 5))
for slicei in range(len(tracklet)):
for bb in tracklet[slicei]:
dcmslice = croppatch(dicomstack[slicei], bb.y, bb.x, bb.h * 2,
bb.w * 2)
w = int(round(bb.w / 2))
h = int(round(bb.h / 2))
dcmslice[h, w:w * 3] = np.max(dcmslice)
dcmslice[h:h * 3, w] = np.max(dcmslice)
dcmslice[h * 3, w:w * 3] = np.max(dcmslice)
dcmslice[h:h * 3, w * 3] = np.max(dcmslice)
plt.subplot(2, int(np.ceil(len(dicomstack) / 2)), slicei + 1)
plt.imshow(dcmslice)
plt.title(slicei)
if figfilename is not None:
plt.savefig(figfilename)
else:
plt.show()
plt.close()
def gen_aug_patch(caseloader, slicei, dcmstack=None):
DEBUG = 0
ctx, cty = caseloader.loadct(slicei)
if dcmstack is None:
dcmstack = caseloader.loadstack(slicei, 'dcm')
cartcont = caseloader.load_cart_cont(slicei)
#cart_patch = croppatch(dcmstack,cty,ctx,256,256)
#plt.imshow(cart_patch)
#plt.show()
mindist = caseloader.load_cart_min_dist(slicei)
dist_pred_ct_gt, nms_dist_map_p_gt = find_nms_center(mindist, fig=1)
validy, validx = np.where(nms_dist_map_p_gt[:, :, 0] > 128)
'''
for rd in range(5):
rnd = random.randint(0,validx.shape[0]-1)
ckey = '%d-%d'%(validx[rnd],validy[rnd])
dist_pred_ct_gt[ckey] = nms_dist_map_p_gt[validy[rnd],validx[rnd],0]/256
'''
xypos = [[validx[i], validy[i]] for i in range(len(validx))]
kmeans = KMeans(n_clusters=5, random_state=0).fit(xypos)
for kmcti in kmeans.cluster_centers_:
kmcx = int(round(kmcti[0]))
kmcy = int(round(kmcti[1]))
ckey = '%d-%d' % (kmcx, kmcy)
if nms_dist_map_p_gt[kmcy, kmcx, 0] < 127:
print('kmean cluster center not large enough', kmcti,
nms_dist_map_p_gt[kmcy, kmcx, 0])
dist_pred_ct_gt[ckey] = nms_dist_map_p_gt[kmcy, kmcx, 0] / 256
#print(dist_pred_ct_gt)
aug_patch_obj = {}
targetdir = DATADESKTOPdir + '/DVWIMAGES/'
aug_cart_patch_batch = []
new_cart_patch_batch_filename = '/casepatch/'+caseloader.art+'/'+caseloader.pjname+'/augcart/'+caseloader.pi+os.path.basename(caseloader.dcmpath(slicei))[:-4]+\
caseloader.side+'.npy'
aug_patch_obj['cartpatchbatchname'] = new_cart_patch_batch_filename
if not os.path.exists(targetdir + '/casepatch/' + caseloader.art + '/' +
caseloader.pjname + '/augcart/'):
os.mkdir(targetdir + '/casepatch/' + caseloader.art + '/' +
caseloader.pjname + '/augcart/')
aug_polar_patch_batch = []
new_polar_patch_batch_filename = '/casepatch/'+caseloader.art+'/'+caseloader.pjname+'/augpolar/'+caseloader.pi+os.path.basename(caseloader.dcmpath(slicei))[:-4]+\
caseloader.side+'.npy'
aug_patch_obj['polarpatchbatchname'] = new_polar_patch_batch_filename
if not os.path.exists(targetdir + '/casepatch/' + caseloader.art + '/' +
caseloader.pjname + '/augpolar/'):
os.mkdir(targetdir + '/casepatch/' + caseloader.art + '/' +
caseloader.pjname + '/augpolar/')
aug_polar_cont_batch = []
new_polar_cont_batch_filename = '/casepatch/'+caseloader.art+'/'+caseloader.pjname+'/augpolarcont/'+caseloader.pi+os.path.basename(caseloader.dcmpath(slicei))[:-4]+\
caseloader.side+'.npy'
aug_patch_obj['polarcoutbatchname'] = new_polar_cont_batch_filename
if not os.path.exists(targetdir + '/casepatch/' + caseloader.art + '/' +
caseloader.pjname + '/augpolarcont/'):
os.mkdir(targetdir + '/casepatch/' + caseloader.art + '/' +
caseloader.pjname + '/augpolarcont/')
aug_patch_obj['auginfo'] = []
#plt.imshow(dcmstack)
#plt.title('dcmstack')
#plt.show()
for cti in dist_pred_ct_gt.keys():
cctx = int(round(int(cti.split('-')[0]) + ctx - 256))
ccty = int(round(int(cti.split('-')[1]) + cty - 256))
#print(cctx,ccty)
tx = cctx - ctx
ty = ccty - cty
new_cart_patch = croppatch(dcmstack, ccty, cctx, 256, 256)
#plt.imshow(new_cart_patch[:,:,1])
#plt.title('new_cart_patch')
#plt.show()
new_cart_patch = new_cart_patch / np.max(new_cart_patch)
aug_cart_patch_batch.append(new_cart_patch)
new_polar_patch = topolar(new_cart_patch, 256, 256)
new_polar_patch = new_polar_patch / np.max(new_polar_patch)
#plt.imshow(new_polar_patch)
#plt.show()
aug_polar_patch_batch.append(new_polar_patch)
#rebase polar cont
new_polar_cont_lumen = tocordpolar(cartcont[0], cctx, ccty)
new_polar_cont_wall = tocordpolar(cartcont[1], cctx, ccty)
aug_polar_cont = np.zeros((256, 2))
aug_polar_cont[:, 0] = new_polar_cont_lumen
aug_polar_cont[:, 1] = new_polar_cont_wall
#exportpolarcontour(new_polar_cont_filename,[new_polar_cont_lumen,new_polar_cont_wall])
aug_polar_cont_batch.append(aug_polar_cont)
if DEBUG:
#check polar ct matches with cart patch
cart_patch_dsp = np.zeros(
(new_cart_patch.shape[0], new_cart_patch.shape[1], 3))
cart_patch_dsp[:, :,
0] = new_cart_patch[:, :,
new_cart_patch.shape[2] // 2]
polarbd = np.concatenate(
[new_polar_cont_lumen[:, None], new_polar_cont_wall[:, None]],
axis=1) * 256
contourin, contourout = toctbd(polarbd, 256, 256)
cart_vw_seg = plotct(512, contourin, contourout)
plt.imshow(cart_vw_seg)
cart_patch_dsp[:, :, 1] = cart_vw_seg
plt.imshow(cart_patch_dsp)
plt.show()
augobj = {}
augobj['ctofx'] = cctx
augobj['ctofy'] = ccty
augobj['transx'] = tx
augobj['transy'] = ty
augobj['augid'] = len(aug_cart_patch_batch) - 1
aug_patch_obj['auginfo'].append(copy.copy(augobj))
#print(augobj)
np.save(targetdir + new_cart_patch_batch_filename,
np.array(aug_cart_patch_batch, dtype=np.float16))
np.save(targetdir + new_polar_patch_batch_filename,
np.array(aug_polar_patch_batch, dtype=np.float16))
np.save(targetdir + new_polar_cont_batch_filename,
np.array(aug_polar_cont_batch, dtype=np.float16))
caseloader.caselist['slices'][slicei]['augpatch'] = aug_patch_obj
with open(caseloader.caselistname, 'wb') as fp:
pickle.dump(caseloader.caselist, fp)
def multi_min_dist_pred_withinbb(min_dist_cnn, dicomslice, bb, patchheight,
patchwidth):
SCALE = 4
octy = bb.y
octx = bb.x
ENL = 1.2
min_dist_search_patch = multi_min_dist_map(min_dist_cnn, dicomslice, octy,
octx, patchheight, patchwidth)
min_dist_search_patch_rz = croppatch(min_dist_search_patch,
min_dist_search_patch.shape[1] / 2,
min_dist_search_patch.shape[0] / 2,
bb.h / 2 * SCALE * ENL,
bb.w / 2 * SCALE * ENL)
min_dist_search_patch_rz = fillpatch(np.zeros(min_dist_search_patch.shape),
min_dist_search_patch_rz)
#plt.imshow(min_dist_search_patch_rz)
#plt.show()
#all cts
dist_pred_ct, nms_dist_map_p = find_nms_center(min_dist_search_patch_rz,
fig=1)
#cts = [[int(i.split('-')[0]), int(i.split('-')[1])] for i in list(dist_pred_ct.keys())]
nms_dist_map = nms_dist_map_p[:, :, 0]
#remove cts in unconnected region
nms_dist_map_int = (nms_dist_map / np.max(nms_dist_map) * 255).astype(
np.uint8)
ret, thresh = cv2.threshold(nms_dist_map_int, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)
connectivity = 4
output = cv2.connectedComponentsWithStats(thresh, connectivity, cv2.CV_32S)
# The second cell is the label matrix
component_label = output[1]
#plt.imshow(component_label)
#plt.show()
center_component_label = component_label[256, 256]
cts = []
for i in dist_pred_ct:
if component_label[int(i.split('-')[1]),
int(i.split('-')[0])] != center_component_label:
#print('ignore unconnected nms ct')
continue
cts.append([int(i.split('-')[0]), int(i.split('-')[1])])
search_patch_center = croppatch(dicomslice, octy, octx, patchheight,
patchwidth)
search_patch_center_rz = cv2.resize(search_patch_center, (0, 0),
fx=4,
fy=4)
# search_patch_center_rz = search_patch_center
nms_dist_map_p[:, :, 2] = search_patch_center_rz / np.max(
search_patch_center_rz) * 255
if len(cts) == 0:
#print('no ct')
pass
elif len(cts) != 1:
pass
'''print('cts', len(cts))
plt.subplot(1,2,1)
plt.title('component_label')
plt.imshow(component_label)
plt.subplot(1, 2, 2)
plt.title('nms_dist_map_p')
plt.imshow(nms_dist_map_p)
plt.show()'''
return cts, nms_dist_map_p
def polar_seg_slice(vwcnn,
vwcfg,
dicomstack,
slicei,
cts,
SCALE=4,
usegrad=False,
ITERPRED=True,
cartreg=False):
DEBUG = 0
contourin = None
contourout = None
polarconsistency = None
recenters = None
if DEBUG:
print(slicei, 'cts', cts)
if len(cts) == 1:
ct = cts[0]
ctx = ct[0]
cty = ct[1]
if ITERPRED == False:
cartstack = crop_dcm_stack(dicomstack, slicei, cty, ctx, 64, 1)
cartstack = cartstack / np.max(cartstack)
polarimg = topolar(cartstack, 64, 64)
polar_img_rz = cv2.resize(polarimg, (0, 0), fx=SCALE, fy=SCALE)
if cartreg:
cart_stack_rz = cv2.resize(cartstack, (0, 0), fx=2, fy=2)
contbd = vwcnn.predict(cart_stack_rz[None, :, :, :,
None])[0] * 256
contour_in_roi = contbd[:, :2] + 256
contour_out_roi = contbd[:, 2:] + 256
contourin = to_dcm_cord(contour_in_roi, BB(ctx, cty, 0, 0))
contourout = to_dcm_cord(contour_out_roi, BB(ctx, cty, 0, 0))
polarconsistency = None
else:
polarbd, polarsd = polar_pred_cont_cst(polar_img_rz[:, :, :,
None],
vwcfg,
vwcnn,
usegrad=usegrad)
polarconsistency = 1 - np.mean(polarsd, axis=0)
# contour positions [x, y] in original dicom space
contourin, contourout = toctbd(polarbd / SCALE, ctx, cty)
if DEBUG:
contourin, contourout = toctbd(polarbd, 256, 256)
cartvw = plotct(512, contourin, contourout)
plt.subplot(1, 2, 1)
cartstackdsp = copy.copy(cartstack)
cartstackdsp[:,
cartstackdsp.shape[1] // 2] = np.max(cartstackdsp)
cartstackdsp[cartstackdsp.shape[0] // 2] = np.max(cartstackdsp)
plt.imshow(cartstackdsp)
plt.subplot(1, 2, 2)
cartvwdsp = copy.copy(cartvw)
cartvwdsp[:, cartvwdsp.shape[1] // 2] = np.max(cartvwdsp)
cartvwdsp[cartvwdsp.shape[0] // 2] = np.max(cartvwdsp)
plt.imshow(cartvwdsp)
plt.show()
# offset in 512 dcm cordinate
else:
# iterative pred and refine center
REPS = 20
MOVEFACTOR = 0.5
MAX_OFF_STEP = 1
mincty = cty
minctx = ctx
mindiff = np.inf
lastoffset = [0, 0]
if DEBUG:
print('init', ctx, cty)
for repi in range(REPS):
if repi == REPS - 1 or MAX_OFF_STEP < 0.1:
if DEBUG:
print('use min diff cts', minctx, mincty,
'with mindiff', mindiff)
cty = mincty
ctx = minctx
cartstack = crop_dcm_stack(dicomstack, slicei, cty, ctx, 64, 1)
cartstack = cartstack / np.max(cartstack)
polarimg = topolar(cartstack, 64, 64)
polar_img_rz = cv2.resize(polarimg, (0, 0), fx=SCALE, fy=SCALE)
polarbd, polarsd = polar_pred_cont_cst(polar_img_rz[:, :, :,
None],
vwcfg,
vwcnn,
usegrad=usegrad)
if DEBUG:
contourin, contourout = toctbd(polarbd, 256, 256)
cartvw = plotct(512, contourin, contourout)
plt.subplot(1, 2, 1)
cartstackdsp = copy.copy(cartstack)
cartstackdsp[:, cartstackdsp.shape[1] //
2] = np.max(cartstackdsp)
cartstackdsp[cartstackdsp.shape[0] //
2] = np.max(cartstackdsp)
plt.imshow(cartstackdsp)
plt.subplot(1, 2, 2)
cartvwdsp = copy.copy(cartvw)
cartvwdsp[:, cartvwdsp.shape[1] // 2] = np.max(cartvwdsp)
cartvwdsp[cartvwdsp.shape[0] // 2] = np.max(cartvwdsp)
plt.imshow(cartvwdsp)
plt.show()
#offset in 512 dcm cordinate
polar_cont_offset = cal_polar_offset(polarbd)
if repi > 0 and MAX_OFF_STEP > 0.1 and lastoffset == polar_cont_offset:
#MAX_OFF_STEP += 0.5/SCALE
MAX_OFF_STEP = 0.09
if DEBUG:
print('move ct no change', MAX_OFF_STEP, lastoffset,
polar_cont_offset)
continue
cofftype = [polar_cont_offset[0] > 0, polar_cont_offset[1] > 0]
if repi == 0:
offtype = cofftype
if repi > 2 and cofftype != offtype:
MAX_OFF_STEP /= 2
if DEBUG:
print('reduce max move to', MAX_OFF_STEP)
offtype = cofftype
polarconsistency = 1 - np.mean(polarsd, axis=0)
#ccstl = polarconsistency[0]
#ccstw = polarconsistency[1]
# contour positions [x, y] in original dicom space
contourin, contourout = toctbd(polarbd / SCALE, ctx, cty)
cdif = np.max(abs(np.array(polar_cont_offset)))
if cdif < 1 or MAX_OFF_STEP < 0.1:
if cdif < 1:
print(polar_cont_offset)
print('==', slicei, 'Break tracklet ref',
polar_cont_offset)
break
if cdif < mindiff:
mindiff = cdif
mincty = cty
minctx = ctx
cofx = polar_cont_offset[0]
cofy = polar_cont_offset[1]
if abs(polar_cont_offset[0]) < 1:
cofx = 0
if abs(polar_cont_offset[1]) < 1:
cofy = 0
if repi < 2:
ctx += cofx * MOVEFACTOR
cty += cofy * MOVEFACTOR
else:
ctx = ctx + max(-MAX_OFF_STEP,
min(MAX_OFF_STEP, cofx * MOVEFACTOR))
cty = cty + max(-MAX_OFF_STEP,
min(MAX_OFF_STEP, cofy * MOVEFACTOR))
print('repeat', repi, 'offset', polar_cont_offset, ctx, cty)
lastoffset = polar_cont_offset
recenters = [[ctx, cty]]
else:
print('multiple cts', cts)
all_contour_in = []
all_contour_out = []
all_polarconsistency = []
for ct in cts:
ctx = ct[0]
cty = ct[1]
cartstack = crop_dcm_stack(dicomstack, slicei, cty, ctx, 64, 1)
cartstack = cartstack / np.max(cartstack)
polarimg = topolar(cartstack, 64, 64)
polar_img_rz = cv2.resize(polarimg, (0, 0), fx=SCALE, fy=SCALE)
polarbd, polarsd = polar_pred_cont_cst(polar_img_rz[:, :, :, None],
vwcfg, vwcnn)
polarconsistency_c = 1 - np.mean(polarsd, axis=0)
contour_in_c, contour_out_c = toctbd(polarbd / SCALE, ctx, cty)
if DEBUG:
plt.subplot(1, 2, 1)
plt.imshow(polarimg)
plt.subplot(1, 2, 2)
polarseg = plotpolar(256, polarbd)
plt.imshow(polarseg)
plt.show()
all_contour_in.append(contour_in_c)
all_contour_out.append(contour_out_c)
all_polarconsistency.append(polarconsistency_c)
contourin = mergecont(all_contour_in, cts)
contourout = mergecont(all_contour_out, cts)
ctx, cty = np.mean(cts, axis=0)
if DEBUG:
seg_vw_merge = plotct(512 * SCALE, contourin * SCALE,
contourout * SCALE)
predseg = croppatch(seg_vw_merge, cty * SCALE, ctx * SCALE,
64 * SCALE, 64 * SCALE)
plt.imshow(predseg)
plt.show()
polarconsistency = np.mean(np.array(all_polarconsistency), axis=0)
#no center adjustment for multi ct prediction
recenters = [[ctx, cty]]
return [contourin, contourout], polarconsistency, recenters
def data_generator(config, exams, aug):
xarray = np.zeros([
config['batchsize'], config['height'], config['width'],
config['depth'], config['channel']
])
yarray = np.zeros([config['batchsize'], config['patchheight'], 2])
bi = 0
while 1:
for ei in exams:
caseloader = CaseLoader(ei)
# print(caseloader)
for slicei in caseloader.slices:
if caseloader.valid_slicei(slicei) == False:
continue
cartcont = caseloader.load_cart_cont(slicei)
min_dist_map = caseloader.load_cart_min_dist(slicei)
dcmstack = caseloader.loadstack(slicei, 'dcm')
ctx, cty = caseloader.loadct(slicei)
aug_polar_patch_batch = np.zeros((10, 256, 256, 3))
aug_polar_cont_batch = np.zeros((10, 256, 2))
offi = 0
mi = 0
while offi < 10 and mi < 50:
mi += 1
if offi == 0:
ofx = 0
ofy = 0
else:
ofx = random.randint(-20, 20)
ofy = random.randint(-20, 20)
cctx = ofx + ctx
ccty = ofy + cty
if min_dist_map[256 + ofx, 256 + ofy] < 0.5:
if mi > 40:
print('mi', mi)
# print('mind',min_dist_map[256+ofx,256+ofy])
continue
# print(ofx,ofy)
new_polar_cont_lumen = tocordpolar(cartcont[0], cctx, ccty)
new_polar_cont_wall = tocordpolar(cartcont[1], cctx, ccty)
aug_polar_cont_batch[offi, :, 0] = new_polar_cont_lumen
aug_polar_cont_batch[offi, :, 1] = new_polar_cont_wall
new_cart_patch = croppatch(dcmstack, ccty, cctx, 256, 256)
new_polar_patch = topolar(new_cart_patch, 256, 256)
aug_polar_patch_batch[offi] = new_polar_patch / np.max(
new_polar_patch)
offi += 1
# aug_polar_patch_batch = caseloader.load_aug_patch(slicei,'polar_patch')
# aug_polar_cont_batch = caseloader.load_aug_patch(slicei,'polar_cont')
for augi in range(len(aug_polar_patch_batch)):
xarray[bi] = aug_polar_patch_batch[augi, :, :, :, None]
yarray[bi] = aug_polar_cont_batch[augi]
bi += 1
if bi == config['batchsize']:
bi = 0
if aug == True:
for offi in random.sample(range(config['height']),
config['rottimes']):
xarray_off = batch_polar_rot(xarray, offi)
yarray_off = batch_polar_rot(yarray, offi)
yield (xarray_off, yarray_off)
else:
yield (xarray, yarray)
def bb_match_score(featuremodel, imgstack, pos1, pos2):
patch1 = croppatch(imgstack[pos1[2]], pos1[1], pos1[0], 64, 64)
patch2 = croppatch(imgstack[pos2[2]], pos2[1], pos2[0], 64, 64)
return bb_match_score_patch(featuremodel, patch1, patch2)