Unet for Medical Image Segmentation using TF 2.x

This example demonstrates the use of U-net model for retinal blood vessel segmentation on tensorflow 2.x

This jupyter notebook presents all requirements needed to achieve pixel-level semantic segmentation using images. This specific use case is on retinal image blood vessel segmentation as explained in the blog series:

https://www.youtube.com/watch?v=1kiPy2tvECs&list=PLQKflBw-kPeecjC345saTF2YrfImciCrr

This specific implementation models the U-net implemenation in the research paper: Roychowdhury S. Few Shot Learning Framework to Reduce Inter-observer Variability in Medical Images. IEEE ICPR, 2021. https://arxiv.org/abs/2008.02952

Step 1: Library requirements

Download the library requirements "requirements.yml" and ensure your python environment is compatible.

Step 2: Train and Test Data

Download the STARE vessel data

• From the page https://cecas.clemson.edu/~ahoover/stare/probing/index.html
• Download 20 images (top row) followed by the labelled vessels by Adam Hoover (second row).
• Unpack the images and labels in a local folder.
• Create the following folder structure in your local folder:

    ./STARE/
            │
            └───train
            |    └───images
            |    └───GT
            │
            └───test
                ├──images
                └───GT

• The following images are used to train: [image0001 image0002 image0003 image0004 image0005 image0044 image0077 image0081 image0082 image0139 ]
• Place these images under ./STARE/train/images/ folder and the hand labelled images under ./STARE/train/GT folder
• All remaining images are placed under ./STARE/test/images/ and the respective hand labelled images are under ./STARE/test/GT/

Step 3: Get the code

Download the U-net_TF2_retinal_image_segmentation_STARE.ipynb and run it.

Step 4: Results:

• The U-net with Depth 4 is trained using 10 images, loss function of binary-crossentropy, Adam optimizer and augmented significnatly using the keras imagedata generator.
• The tensorboard graphs are as follows:

How to avoid Errors in the Code?

The key component of this U-net framework is that input is an image and output is also an image. To use the code in this repo AS IS, you HAVE TO unpack the data set as suggested in Step 2.

• One way to detect if your Path is incorrectly set is, you will get the message: "Found 0 images beloning to 1 classes". This means the images and GT are not detected.
• If images are not detected, this will lead to a "peek" error while model.fit command is run.
• Finally, for tensorboard, if no visualization is created, check the 'logs' folder. If a recored exists then rerun the tensorboard command. This should get things started.

The segmentation perfromances on test images are:

Precision= 0.7557, Recall= 0.8493 IoU= 0.6648 acc= 0.9606 F1= 0.7979

• Sample examples are: Predicted output:
• Other outputs under STARE(Copy)/test/pred/