U-Net | simple U-Net implementation | Machine Learning library

I encountered many hardships when trying to fit a CNN (U-Net) to my tif training images in Python.

I have the following structure to my data:

• X
• • 0
• • • [Images] (tif, 3-band, 128x128, values ∈ [0, 255])
• X_val
• • 0
• • • [Images] (tif, 3-band, 128x128, values ∈ [0, 255])
• y
• • 0
• • • [Images] (tif, 1-band, 128x128, values ∈ [0, 255])
• y_val
• • 0
• • • [Images] (tif, 1-band, 128x128, values ∈ [0, 255])

Starting with this data, I defined ImageDataGenerators:

I am trying to segment medical images using a version of U-Net implemented with Keras. The inputs of my network are 3D images and the outputs are two one-hot-encoded 3D segmentation maps. I know that my dataset is very imbalanced (there is not so much to segment) and therefore I want to use class weights for my loss function (currently binary_crossentropy). With the class weights, I hope the model will give more attention to the small stuff it has to segment.

If you know the imbalance of your database, you can pass the parameter class_weight to model.fit(). Does this also work with my use case?

I am coming from medical background and a newbie in this machine learning field. I am trying to train my U-Net model using keras and tensorflow for image segmentation. However, my loss value is all NaN and the prediction is all black.

I would like to check the U-Net layer by layer but I don't know how to feed the data and from where to start. What I meant by checking for each layer is that I want to feed my images to first layer for example and see the output from the first layer and then moving on to the second layer and until to the last layer. Just want to see how the output is produced for each layer and to check from where the nan value is started. Really appreciate for your help.

These are my codes.

I am using the Image segmentation guide by fchollet to perform semantic segmentation. I have attempted modifying the guide to suit my dataset by labelling the 8-bit img mask values into 1 and 2 like in the Oxford Pets dataset which will be subtracted to 0 and 1 in class Generator(keras.utils.Sequence).The input image is an RGB-image.

I am not sure why but my dice coefficient isn't increasing at all. I have tried to reduce the learning rate as well as changing the optimizer to SGD/RMSProp, normalizing the data, taking the imbalanced labels into account but the result is very strange. The accuracy/IoU of the model is decreasing as the no. of epochs increases.

If it helps, I previously asked a question about the metrics that I should be using for an imbalanced dataset here. The visualization of the predictions are okay but the metric is not.

What I can do next to debug this problem? Is there anything wrong with my code? Will appreciate any advice.

Here are the results

I am trying to train a model (U-Net) on RGB images with shape of ( 256, 256, 3) but when I fit the model I get the following error:

I have a code which gives me binary images using Otsu thresholding. I am making a dataset for a U-Net, and I want to try different algorithms (global as well as local) for the same, so that I can save the "best" image. Below is the code for my image binarization.

I've defined a U-Net model using Pytorch but it won't accept my input. I've checked the model layers and they seem to be applying the operations as I would expect them to but I still get an error.

I've just switched to Pytorch after mostly using Keras so I'm not really sure how to debug this issue, the error I get is:

RuntimeError: Given groups=1, weight of size [32, 64, 3, 3], expected input[1, 128, 65, 65] to have 64 channels, but got 128 channels instead

Here's the code I'm using:

I'm implementing a U-Net based architecture in PyTorch. At train time, I've patches of size 256x256 which doesn't cause any problem. However at test time, I've full HD images (1920x1080). This is causing a problem during skip connections.

Downsampling 1920x1080 3 times gives 240x135. If I downsample one more time, the resolution becomes 120x68 which when upsampled gives 240x136. Now, I cannot concatenate these two feature maps. How can I solve this?

PS: I thought this is a fairly common problem, but I didn't get any solution or even mentioning of this problem anywhere on the web. Am I missing something?

I'm training a network with MRI images and I wanted to use SSIM as loss function. Till now I was using MSE, and everything was working fine. But when I tried to use SSIM (tf.image.ssim), I get a bunch of these warining messages: