def report_scores(self, y, t):
with chainer.no_backprop_mode():
if self.nested_label:
dice = mean_dice_coefficient(
dice_coefficient(y[:, 0:2, ...], t[:, 0, ...]))
for i in range(1, t.shape[1]):
dices = dice_coefficient(y[:, 2 * i:2 * (i + 1), ...],
t[:, i, ...])
dice = F.concat((dice, mean_dice_coefficient(dices)),
axis=0)
else:
dice = dice_coefficient(y, t)
mean_dice = mean_dice_coefficient(dice)
if self.nested_label:
b, c, h, w, d = t.shape
y = F.reshape(y, (b, 2, h * c, w, d))
t = F.reshape(t, (b, h * c, w, d))
accuracy = F.accuracy(y, t)
reporter.report({'acc': accuracy, 'mean_dc': mean_dice})
xp = cuda.get_array_module(y)
for i in range(len(dice)):
if not xp.isnan(dice.data[i]):
reporter.report({'dc_{}'.format(i): dice[i]})
def compute_dice_coef(self, y, t):
if self.nested_label:
dice = mean_dice_coefficient(dice_coefficient(y[:, 0:2, ...], t[:, 0, ...]))
for i in range(1, t.shape[1]):
dices = dice_coefficient(y[:, 2*i:2*(i+1), ...], t[:, i, ...])
dice = F.concat((dice, mean_dice_coefficient(dices)), axis=0)
else:
dice = dice_coefficient(y, t, is_brats=self.is_brats)
return dice
def evaluate(self):
summary = reporter_module.DictSummary()
iterator = self._iterators['main']
enc = self._targets['enc']
dec = self._targets['dec']
bound = self._targets['bound']
reporter = Reporter()
observer = object()
reporter.add_observer(self.default_name, observer)
if hasattr(iterator, 'reset'):
iterator.reset()
it = iterator
else:
it = copy.copy(iterator)
for batch in it:
x, t, te = self.converter(batch, self.device)
with chainer.no_backprop_mode(), chainer.using_config(
'train', False):
if self.dataset == 'msd_bound':
h, w, d = x.shape[2:]
hc, wc, dc = self.crop_size
if self.nested_label:
y = cupy.zeros((1, 2 * (self.nb_labels - 1), h, w, d),
dtype='float32')
else:
y = cupy.zeros((1, self.nb_labels, h, w, d),
dtype='float32')
s = 128 # stride
ker = 256 # kernel size
dker = dc # kernel size for depth
ds = dker * 0.5 # stride for depth
dsteps = int(math.floor((d - dker) / ds) + 1)
steps = round((h - ker) / s + 1)
for i in range(steps):
for j in range(steps):
for k in range(dsteps):
xx = x[:, :, s * i:ker + s * i,
s * j:ker + s * j, ds * k:dker + ds * k]
hhs = enc(xx)
yye, bbs = bound(hhs)
yy = dec(hhs, bbs)
y[:, :, s * i:ker + s * i, s * j:ker + s * j,
ds * k:dker + ds * k] += yy.data
# for the bottom depth part of the image
xx = x[:, :, s * i:ker + s * i, s * j:ker + s * j,
-dker:]
hhs = enc(xx)
yye, bbs = bound(hhs)
yy = dec(hhs, bbs)
y[:, :, s * i:ker + s * i, s * j:ker + s * j,
-dker:] += yy.data
else:
hs = enc(x)
ye, bs = bound(hs)
y = dec(hs, bs)
seg_loss = self.compute_loss(y, t)
accuracy = self.compute_accuracy(y, t)
dice = self.compute_dice_coef(y, t)
mean_dice = mean_dice_coefficient(dice)
weighted_loss = seg_loss
observation = {}
with reporter.scope(observation):
reporter.report(
{
'loss/seg': seg_loss,
'loss/total': weighted_loss,
'acc': accuracy,
'mean_dc': mean_dice
}, observer)
xp = cuda.get_array_module(y)
for i in range(len(dice)):
if not xp.isnan(dice.data[i]):
reporter.report({'dc_{}'.format(i): dice[i]},
observer)
summary.add(observation)
return summary.compute_mean()
def evaluate(self):
summary = reporter_module.DictSummary()
iterator = self._iterators['main']
enc = self._targets['enc']
dec = self._targets['dec']
reporter = Reporter()
observer = object()
reporter.add_observer(self.default_name, observer)
if hasattr(iterator, 'reset'):
iterator.reset()
it = iterator
else:
it = copy.copy(iterator)
for batch in it:
x, t = self.converter(batch, self.device)
with chainer.no_backprop_mode(), chainer.using_config('train', False):
if self.dataset == 'msd_bound':
# evaluation method for BRATS dataset only
h, w, d = x.shape[2:]
hc, wc, dc = self.crop_size
if self.nested_label:
y = cupy.zeros((1, 2*(self.nb_labels-1), h, w, d), dtype='float32')
else:
y = cupy.zeros((1, self.nb_labels, h, w, d), dtype='float32')
hker = hc # kernel size
hs = int(0.5*hker) # stride
wker = wc
wc = int(0.5*wker)
dker = dc # kernel size for depth
for i in range(2):
for j in range(2):
for k in range(2):
xx = x[:, :, -i*hker:min(hker*(i+1), h),
-j*wker:min(wker*(j+1), w), -k*dker:min(dker*(k+1), d)]
hs = enc(xx)
yy = dec(hs)
y[:, :, -i*hker:min(hker*(i+1), h),
-j*wker:min(wker*(j+1), w),
-k*dker:min(dker*(k+1), d)] += yy.data
else:
hs = enc(x)
y = dec(hs)
seg_loss = self.compute_loss(y, t)
accuracy = self.compute_accuracy(y, t)
dice = self.compute_dice_coef(y, t)
mean_dice = mean_dice_coefficient(dice)
observation = {}
with reporter.scope(observation):
reporter.report({
'loss/seg': seg_loss,
'acc': accuracy,
'mean_dc': mean_dice
}, observer)
xp = cuda.get_array_module(y)
for i in range(len(dice)):
if not xp.isnan(dice.data[i]):
reporter.report({'dc_{}'.format(i): dice[i]}, observer)
summary.add(observation)
return summary.compute_mean()
