add support for UNET++

[michalhg]

UNET++ ustilizes skip connections to exploit multi-scale features for image segmentation. it added dense skip connection adn deep supervision to the original UNET
https://arxiv.org/pdf/1912.05074.pdf

please add this network to the networks file.
official keras implementation is given in : https://github.com/MrGiovanni/UNetPlusPlus

other pytorch implemetation is also given in : https://towardsdatascience.com/biomedical-image-segmentation-unet-991d075a3a4b

[Nic-Ma]

Hi @michalhg ,

Thanks for your feedback.
I marked this request for working group discussion.

Thanks.

[yashika-git]

Hi @Nic-Ma,

I would love to add UNet++ to MONAI.
Though I have never contributed code in open-source, I'm excited to get started.
Will it be possible for you to guide me a little if I get stuck somewhere?

Thank You

[Nic-Ma]

Hi @yashika-git ,

Of course, please go ahead to submit a PR, we can work together to address issues later.
You can refer to our contribution guide:
https://github.com/Project-MONAI/MONAI/blob/dev/CONTRIBUTING.md

Thanks for your contribution!

[yashika-git]

@Nic-Ma,
Sure, I'm at it.
Read about UNET++ yesterday, need sometime to understand it properly in order to implement.

[yashika-git]

Hi @Nic-Ma,

I was going through some codes in order to get an idea of how to implement UNet++.
I had a doubt, in the below line of the file convolutions.py:

MONAI/monai/networks/blocks/convolutions.py

Line 129 in 0b97bb6

if padding is None:

If the user input for padding is None, a same_padding is applied.
I was wondering if user should be given an option to also apply a valid_padding (or no padding) ?

Thank You

[Nic-Ma]

Hi @ericspod ,

Could you please help share some suggestions here? I think you are the expert on the UNet implementation.

Thanks in advance.

[ericspod]

You can provide a padding value of 0 to not pad or some other value calculated as a valid pad value.

[yashika-git]

Thank you for the information, @ericspod.

[baibizhe]

Thank you for the information, @ericspod.

Hello .I would really like to do contibution on this issues. Is there anything that I could do about this issue? I am not familiar with MONAI , but I am familiar with pytorch,python and UNET. If you need any help, please contact me .

[wyli]

Hi @baibizhe, thanks for your interest here. There's a basic UNet implementation in MONAI, the goal was to create an easy-to-understand baseline but with some price in flexibility

MONAI/monai/networks/nets/basic_unet.py

Line 189 in 78509ae

class BasicUNet(nn.Module):

as you can see the forward function is straightforward with a hard-coded structure

MONAI/monai/networks/nets/basic_unet.py

Lines 286 to 300 in 78509ae

	x0 = self.conv_0(x)
	x1 = self.down_1(x0)
	x2 = self.down_2(x1)
	x3 = self.down_3(x2)
	x4 = self.down_4(x3)
	u4 = self.upcat_4(x4, x3)
	u3 = self.upcat_3(u4, x2)
	u2 = self.upcat_2(u3, x1)
	u1 = self.upcat_1(u2, x0)
	logits = self.final_conv(u1)
	return logits

but the activations/dropout are quite flexible using the MONAI APIs:

MONAI/monai/networks/nets/basic_unet.py

Lines 199 to 203 in 78509ae

	act: Union[str, tuple] = ("LeakyReLU", {"negative_slope": 0.1, "inplace": True}),
	norm: Union[str, tuple] = ("instance", {"affine": True}),
	bias: bool = True,
	dropout: Union[float, tuple] = 0.0,
	upsample: str = "deconv",

A unet++ implementation following a similar style (emphasising code readability) would also be quite helpful for the community. is this something that you could contribute? thanks!

[yashika-git]

Thank you for sharing this @wyli.
Can I and @baibizhe work together on this issue?