New loss function for segmentation

[fepegar]

Hi all,

I have implemented the Tversky loss function for segmentation with imbalanced data. I'll try it tomorrow and if it works I'll open a PR.

It needs two parameters α and β to weight precision and recall during training (if α + β = 1, it's equivalent to an Fβ-score and if α = β = 0.5, it's Dice). Is there a way to specify those without hard-coding them? The only other losses that accept kwargs are GDSC and wasserstein_disagreement_map, but I the only example I've found in the repo in which those are passed is in this test:

NiftyNet/tests/loss_segmentation_test.py

Lines 94 to 110 in 0bdbd9c

	def test_generalised_dice_score_uniform_regression(self):
	with self.test_session():
	predicted = tf.constant([[0, 10], [10, 0], [10, 0], [10, 0]],
	dtype=tf.float32, name='predicted')
	labels = tf.constant([[1, 0, 0, 0]], dtype=tf.int64, name='labels')
	weights = tf.cast(labels, tf.float32)
	predicted, labels, weights = [tf.expand_dims(x, axis=0) for x in
	(predicted, labels, weights)]
	test_loss_func = LossFunction(
	2, loss_type='GDSC',
	loss_func_params={'type_weight': 'Uniform'})
	one_minus_generalised_dice_score = test_loss_func(
	predicted, labels, weights)
	self.assertAllClose(one_minus_generalised_dice_score.eval(),
	0.3333, atol=1e-4)

[wyli]

Thanks @fepegar, the loss function's interface is ready, but we need the new configuration file #166 to support a proper nested parameter parsing.

Perhaps @BostonChildrensRadiology would be interested in this implementation as well.

[fepegar]

Cool, looking forward to the new config format.

I'll submit the PR as a WIP in a second.

[amartya-k]

Is this working for class imbalanced semantic segmentation?(mostly medical images)

[fepegar]

@amartya-k yes, it should. Take a look at the paper I referenced in the first comment.