Thank you very much for subscribing and also spreading the word. I really appreciate it.

BRATS is on my list but I need to find time to work it out. I will try. Thanks for the suggestion.

very helpfull sir

Thanks a lot!

You are welcome!

Hi #NOUR-EDDINE did you find how to obtain masks

​@@zakariasaidi2191 using shapefiles + satellite images in arcgis

@NOUR-EDDINE ELOUAFI please I would love to know how you created the masks in ArcGIS. I am currently working on something similar.. I have tried the "Exporting training dataset for deep learning" in ArcGIS pro, but the label image I have is just black image. Please I'm am really in need of this. Looking forward to hear from you soon.

@@temitopeakinboyewa3137 You need to adjust the contrast from 255 to class number. You can use Imagej for that

@@faaalsh8784 Thanks so much. I will try this.

Because our prediction is binary, either background or cell. Our ground truth training masks have binary values (0 and 255), therefore it makes sense to limit our prediction to binary. Please go ahead and change bool to uint8 to see how the output looks.

@@DigitalSreeni Sir why loss become negative when we change Y_train dtype and how we can use float dtype instead of bool and getting a positive result.

If all your images and masks are 256x256 then change the input to architecture to 256. Have a look at the model summary to make sure it makes sense. Another option you have is to cut down your 256x256 images as 4 patches of 128x128 each and use them for training.

@@DigitalSreeni awesome, thank you so much. Greetings from Germany.

me too.. same problem T.T

I dont undestand this part of the code. What is the meaning of multiplying by 0.9 the with the 128? because X_train.shape[0] == 128?

did you find out how to do that?

Depends on what you are comparing. Image segmentation refers to pixel classification. Therefore, accuracy can be simply a measure between original label vs predicted label at each pixel. But pixel accuracy does not tell you much about imbalanced classes. Therefore, intersection over union may be a better metric. Also look up Dice coefficient.

@@DigitalSreeni Thank you for your reply sir. I will read more about what you have mentioned!

mask_ = imread(path/to your/mask/file)[:,:] mask = np.expand_dims(resize(mask_, (IMG_HEIGHT, IMG_WIDTH), mode='constant',preserve_range=True), axis=-1) Y_train[n] = mask

having the exact same issue here

fixed this by fixing train_path it was not set properly to 'stage1_train'

@@amiralessigue9075 Hello amira, how did you fix this? Its happening here too and I don't know what to do

@@marcusbranch2100 May be a little late, but I fixed it by seting the aboslute path to the variables TRAIN_PATH and TEST_PATH... a momentary solution until I find why relative path doesn't work with os.walk...

I had the same problem, after I change the stage1_train and stage1_test to in my computer folder way it was fixed. I hope it will be help all of us. Thanks.

In the data set I've used for the video, we have a total 670 images for training. Of those I used 90% (603) for prediction which is where I used the term - X_train[:int(X_train.shape[0]*0.9)]. This just returns the array of 608 images, each 128x128x3. Similarly, X_train[int(X_train.shape[0]*0.9):] returns the remaining 10% data which I assigned to a variable preds_val.

I am also unsure of this

You can plot train image if you want. I assume if it works for test, it also works for train.

I have done many videos on using custom data set, please check other videos on my channel.

Since my example performs binary segmentation I encoded my Y values as binary. For multi-class segmentation you can define them as integers 0, 1, 2, .... I've also seen examples where people used colors for each class, for example red (255,0,0), blue (0, 255, 0), green (0, 0, 255).

Nah how to do that sir? I still confused how ti make itu 3 classes

How to define it :(

@@adithiajovandy8572 could you find a solution for this? I have the same problem

Yes, of course. You can use U-net to segment fat and tissue as long as a human can segment them. If even a human cannot tell the difference then the machine has no chance.

did you find out how?

@@SawsanAAlowa yeah I used the label me but only for single object segmentation

@@SawsanAAlowa kzfaq.info/get/bejne/q8qCabao1NC9iKM.html I followed this playlist

@@bhanusrinivaskoppolu4814 Hi, Thank you for the fast response. does label me provide converting from JSON to MASK as well or only produces the JSON file.

@@SawsanAAlowa yeah I later converted the file to png format and used it

The system doesn't care if it is a microscope image or a satellite image.

Not as easy as changing one little thing. Please search for 3D unet topics on KZfaq so you can do it properly. Unfortunately I do not have any example code to share.

@@DigitalSreeni Thank you.

This code is unique to the images used in the video. In this case the masks are such a way each object is represented in a separate image. So if a training image has 100 objects then we have 100 images for masks, each showing a separate object. Therefore, I had to merge all masks into single array. If your mask has all objects your job is easy. Ultimately you know your images better so get them in the right format to comply with Unet input requirements.

@@DigitalSreeni sir, why it you took "[2]" for mask_file in next(os.walk(path + '/masks/'))[2]

Of course, if you have multiple regions for segmentation then the labeled masks would be multinary (not sure if that is a word). In other words, your masks will have pixel values 0, 1, 2, 3, etc. depending on the number of classes. For binary the pixel values would be wither 0 or 1.

@@DigitalSreeni so we need to have a ground truth with different pixel value not a binary image is that what you mean

Just pre-process all images like you normally would do. Read images one by one and apply preprocessing operations such as smoothing and then add them to the dataframe that goes in as input to the neural network. A simple for loop would do the job. Saying that, I do not recommend such operations for deep learning as noise often contains information about your image. If you clean it up or do any image processing operations it may not help and potentially could hurt the results. Of course, it depends on the application but just be careful with any such operations.

@@DigitalSreeni OK Thank You Sir...it was of great help.

Glad it was helpful!

@@DigitalSreeni Hi, What do we do if we don't have masks? Can we create these?

Please watch my videos 204 and later for more U-net related videos. For example, this one talks about multiclass U-net. kzfaq.info/get/bejne/jt-IaKt_2tyRqHk.html

Use nibabel library to load nii images. Using this library you can real nii images as numpy arrays and then this approach can be used. I will be releasing new videos on the topic of Brats dataset segmentation so please stay tuned. Also, I hope you watched my latest videos on U-net.

@@DigitalSreeni Just Saw, I was looking for this 3D-Unet. You are awesome :)

next will throw that error if it sees no directories to walk through. Try using a for loop instead.... for path, dirs, files in os.walk('your path to directory'): Print(path) Print(dirs) Print(files)

or try: try: train_ids = next(os.walk(... except StopIteration: pass

Can u give exact code to fix that please

​@@DigitalSreenisir please help me I'm arising same problem what is exact code for it

It has been a while since I wrote that code but upon quick look here is what I can say... At this part of the code: mask = np.maximum(mask, mask_) we are taking maximum value at a given pixel between mask and mask_ images. The mask image has a dimension of (128, 128, 1) and mask_ originally has a dimension of (128, 128) so in order to make the dimensions same between both images we add (extend) dimensions of the mask_ image.

@@DigitalSreeni This helps a lot! Thank you :)

For supervised segmentation you need to provide ground truth for segmentation, for example in the form of masks / labels. You can perform unsupervised segmentation where the system can segment without ground truth; for example using k-means or GMM. You can also perform manual segmentation using image histogram and other image processing operations. The type of image you have and your segmentation needs defines the approach.

@@DigitalSreeni thank you so much

Not sure what you are trying to do. Why would you only have images and not masks for semantic segmentation using Unet? You need to have masks, otherwise how would the network know what to model your images to?

@@DigitalSreeni how can we create the masks. please advise.

@@DigitalSreeni for image reconstruction maybe

You're welcome.

You can use dice coefficient if you want. Normally people use cross entropy because everyone uses it. I am not proficient with dice coefficient but cross entropy has gradients nice to work with back propagation. Please check with any online examples where people used dice; if it isn't common there must be a reason.

@@DigitalSreeni thanks for your answer sir, dice coeff work only if we change dtype of Y_train from bool to uint or float im not sure but i think crossntroy it's used for binary segmentation ..,what i'm doing it's segmentation of multiple région so when i try to change dtype of Y_train and use dice coeff it's give me bad results. Thanks again for those helpful information.

it give me a negative loss

Y-train is Boolean in my example as it is binary segmentation. You can use other values. In fact, for multiclass segmentation it is customary to assign y labels as 0, 1, 2, 3, .... I’m not sure why it isn’t working for you but technically it is not U-net limitation.

@@DigitalSreeni thanks for your reply, i apply the same code for segmentation of alzheimer disease dataset but not give me results . i can't know the reason .please if you want to help me contact me please .

Then you do not use U-Net as it is used for semantic segmentation. You seem to be looking for classification as you'd like to classify an entire image.

@@DigitalSreeni thank you for replying, i am developing two approaches and in 1st approach Entire image is used for classification and for the second approach I have to segment the region of interest and then classification will be done. I want to Use U-Net for all these.

@@DigitalSreeni I will try to Mask image as it is required to provide ground truth

Please look into Mask RCNN approach, may be it is fit for your needs!

@@DigitalSreeni sure! I will go through it, thank you..

Not sure what BRATS data set is? Is it Brain Tumor related? Can you share a link to the data set?

@@DigitalSreeni Sorry Sir for being late to reply, due to pandemic issues. yes, it brain tumor segmentation dataset. Here is the link www.med.upenn.edu/cbica/brats2019/data.html

One more question, what are preprocessing kind of jobs, we can apply to the 2018 data science bowl dataset.

You can find code here: github.com/bnsreenu/python_for_microscopists

Not sure how to help as errors seems ambiguous. Did you try running my code as-is? You can find my code on my GitHub page, github.com/bnsreenu/python_for_microscopists

Stop iteration occurs when you have not set the training path to stage1_train properly..

@@DigitalSreeni same problem here

@@DigitalSreeni same problem here

same problem here

A great question!!!! I failed to find a good solution for mask creation. We are developing a solution for that at my work and it will be released soon. Please sign up at www.apeer.com and keep on eye on the upcoming updates. APEER is a free online platform for image processing that you can use using a browser on any device. In any case, we will hopefully have a solution for you in a few weeks.

you can use imagej software for manual masking

I'm not sure what you mean. Please watch parts 1 to 4 videos for other details. You can also find my code here: github.com/bnsreenu/python_for_microscopists

@@DigitalSreeni Thanks for responding. I meant for this Unet implementation can you provide the version numbers of all packages !! Like Keras==2.0.0, matplotlib==3.01 likewise... all the packages' version numbers. Anyway thanks again :)

@@shahnewazjubayed8159 Oh, I get it. I will have to check, but good point. I work with 3 different systems so I need to find out which one I used for what tutorial. I know, so irresponsible of me but we all learn as we make mistakes :)

did you solve it ? I am having same problem

I'm seeing exact error did u fix it