def test_smooth(
self,
value: float,
smooth_nr: float,
smooth_dr: float,
expected: float,
):
"""
Test values in extreme cases where numerator/denominator are all zero.
:param value: value for input.
:param smooth_nr: constant for numerator.
:param smooth_dr: constant for denominator.
:param expected: target value.
"""
shape = (1, 10)
y_true = tf.ones(shape=shape) * value
y_pred = tf.ones(shape=shape) * value
got = label.DiceScore(smooth_nr=smooth_nr, smooth_dr=smooth_dr).call(
y_true,
y_pred,
)
expected = tf.constant(expected)
assert is_equal_tf(got[0], expected)
def test_background_weight_err(self, background_weight: float):
"""
Test the error message when using wrong background weight.
:param background_weight: weight for background class.
"""
with pytest.raises(ValueError) as err_info:
label.DiceScore(background_weight=background_weight)
assert "The background weight for Dice Score must be within [0, 1]" in str(
err_info.value)
def test_get_config(self):
got = label.DiceScore().get_config()
expected = dict(
binary=False,
neg_weight=0.0,
scales=None,
kernel="gaussian",
reduction=tf.keras.losses.Reduction.SUM,
name="DiceScore",
)
assert got == expected
def test_call(self, y_true, y_pred, binary, neg_weight, scales, expected):
expected = np.array([expected] *
self.shape[0]) # call returns (batch, )
got = label.DiceScore(binary=binary,
neg_weight=neg_weight,
scales=scales).call(y_true=y_true, y_pred=y_pred)
assert is_equal_tf(got, expected)
got = label.DiceLoss(binary=binary,
neg_weight=neg_weight,
scales=scales).call(y_true=y_true, y_pred=y_pred)
assert is_equal_tf(got, -expected)
def test_get_config(self):
got = label.DiceScore().get_config()
expected = dict(
binary=False,
background_weight=0.0,
smooth_nr=1e-5,
smooth_dr=1e-5,
reduction=tf.keras.losses.Reduction.AUTO,
name="DiceScore",
)
assert got == expected
def calculate_metrics(
fixed_image: tf.Tensor,
fixed_label: (tf.Tensor, None),
pred_fixed_image: (tf.Tensor, None),
pred_fixed_label: (tf.Tensor, None),
fixed_grid_ref: tf.Tensor,
sample_index: int,
) -> dict:
"""
Calculate image/label based metrics.
:param fixed_image: shape=(batch, f_dim1, f_dim2, f_dim3)
:param fixed_label: shape=(batch, f_dim1, f_dim2, f_dim3) or None
:param pred_fixed_image: shape=(batch, f_dim1, f_dim2, f_dim3)
:param pred_fixed_label: shape=(batch, f_dim1, f_dim2, f_dim3) or None
:param fixed_grid_ref: shape=(1, f_dim1, f_dim2, f_dim3, 3)
:param sample_index: int,
:return: dictionary of metrics
"""
if pred_fixed_image is not None:
y_true = fixed_image[sample_index:(sample_index + 1), :, :, :]
y_pred = pred_fixed_image[sample_index:(sample_index + 1), :, :, :]
y_true = tf.expand_dims(y_true, axis=4)
y_pred = tf.expand_dims(y_pred, axis=4)
ssd = image_loss.SumSquaredDifference()(y_true=y_true,
y_pred=y_pred).numpy()
else:
ssd = None
if fixed_label is not None and pred_fixed_label is not None:
y_true = fixed_label[sample_index:(sample_index + 1), :, :, :]
y_pred = pred_fixed_label[sample_index:(sample_index + 1), :, :, :]
dice = label_loss.DiceScore(binary=True)(y_true=y_true,
y_pred=y_pred).numpy()
tre = label_loss.compute_centroid_distance(
y_true=y_true, y_pred=y_pred,
grid=fixed_grid_ref[0, :, :, :, :]).numpy()[0]
else:
dice = None
tre = None
return dict(image_ssd=ssd, label_binary_dice=dice, label_tre=tre)
def test_exact_value(self, binary: bool, background_weight: float,
shape: Tuple):
"""
Test dice score by comparing at ground truth values.
:param binary: if project labels to binary values.
:param background_weight: the weight of background class.
:param shape: shape of input.
"""
# init
shape = (1, ) + shape # add batch axis
foreground_weight = 1 - background_weight
tf.random.set_seed(0)
y_true = tf.random.uniform(shape=shape)
y_pred = tf.random.uniform(shape=shape)
# obtained value
got = label.DiceScore(
binary=binary,
background_weight=background_weight,
).call(y_true=y_true, y_pred=y_pred)
# expected value
flatten = tf.keras.layers.Flatten()
y_true = flatten(y_true)
y_pred = flatten(y_pred)
if binary:
y_true = tf.cast(y_true >= 0.5, dtype=y_true.dtype)
y_pred = tf.cast(y_pred >= 0.5, dtype=y_pred.dtype)
num = foreground_weight * tf.reduce_sum(
y_true * y_pred, axis=1) + background_weight * tf.reduce_sum(
(1 - y_true) * (1 - y_pred), axis=1)
num *= 2
denom = foreground_weight * tf.reduce_sum(
y_true + y_pred, axis=1) + background_weight * tf.reduce_sum(
(1 - y_true) + (1 - y_pred), axis=1)
expected = (num + EPS) / (denom + EPS)
assert is_equal_tf(got, expected)
