hamidriasat/unet-3-plus

A Full-Scale Connected UNet for Medical Image Segmentation

unet-3-plus's Issues

cls are you use in bce loss

Hi author, did you use binary cross entropy for supervised classification in the loss, I noticed this is mentioned in the author's paper.

error :batch_images[i] = image ValueError: setting an array element with a sequence.

File "/home/whj2023/unet3/unet3/data_generator.py", line 79, in getitem
return self.__data_generation(indexes)
File "/home/whj2023/unet3/unet3/data_generator.py", line 144, in __data_generation
batch_images[i] = image
ValueError: setting an array element with a sequence.

why?

UNET3+ with DS and CGM

Hi,

May I ask which part of the code should I change to make unet3+ with DS and CGM trainable? I got INVALID_ARGUMENT: required broadcastable shapes error when trying to train with it.

Specifically, I am only using your code to initialize the model, hopefully, I want it to be trained with the input shape of (None, None,8) for binary segmentation. Please let me know if you want more details.

Best
Yili

About hybrid loss

Hi, sorry for disturbing you. Do you have any pytorch code for hybrid loss you have used here? if so, I would be really grateful if you share it with me.

tiny little typos

This code reads very well. It's nicely coded and easy to read and understand the whole idea of the U-Net3+. So many thanks to you.

I just want to let you know that there might be some typos.

d4_d3 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(d4)
d4_d3 = conv_block(d4_d3, cat_channels, n=1)

I guess all 'd4's should be 'e4'
e4_d3 = k.layers.UpSampling2D(size=(2, 2), interpolation='bilinear')(e4)
e4_d3 = conv_block(e4_d3, cat_channels, n=1)

I couldn't understand your logic here.

Hi;

I'm not trying to point to any issue. I just couldn't understand your logic. My understanding is in any Unet you need both the output of pooling (for downward path) and output of convolution for the skip connection. so some thing like this:

c1, p1 = down_block(p0, f[1])

But I notice in your code your are overwriting the skip value, like this:

e2 = k.layers.MaxPool2D(pool_size=(2, 2))(e1)  # 160*160*64
e2 = conv_block(e2, filters[1])                    # 160*160*128

here e1 gets overwritten, which confuse me. I hope my question is clear.

Thank you!

Just 2 questions

Hi,

Thanks for making the code of UNet-3-Plus simpler, however I have 2 questions and I would be very grateful to hear from you:

1. In the paper the author says they used ReseNet-101 as backbone encoder, but in your code I could not find where exactly ReseNet-101is used? is it used at all? if not how did you extract the features? using custom convolutional layers?
2. My second question is about the decoder's logic, did you use custom functions for Max pooling, Relu, batch normalization... etc?