Brain-Tumor-segmentation-using-UNET-AND-RESUNET

This notebook demonstrates metrics for segmentation and performance of 2 different models for segmentation on a kaggle dataset

Link to the trained models

RES-Unet Model

https://drive.google.com/file/d/13Q3IOPHDLzvqo0KMD3vw-kLMjY7y5OTW/view?usp=sharing

UNET model

https://drive.google.com/file/d/1RXJXcsFJqBUi3FrCgRY2UsBQ04tJXSJk/view?usp=sharing

INPUT Requirments

• The input has to be grayscale image of shape 256*256*1
• The intensities have to be normalize by dividing them by 255.0

Brain tumor segmentation with RESUNET/UNET

Using UNET to determine brain
tumors
By-
Shivansh Singla
UIET CSE

Explaining the problem

We are given various images of brain scans and brain tumor masks corresponding to it.Our goal is to train our model such that given an image of brain it should be able to detect the tumor.

Dataset description

Our dataset has 4 folders

1. Images contain the images of brain scans.
2. Masks contain the segmented tumors which we need to detect
3. Saliency maps 1
4. Saliency maps 2
5. Approx size of dataset 900MB
6. Thera are 3064 images and masks

What Is a Mask?

The goal of image segmentation is to understand image at the pixel level. It associates each pixel
with a certain class. The output produced by image segmentation model is called MASK of the
image.

Mask can be represented by associating pixel values with their coordinates for example it we have a

black image of shape (2,2) this can be represented as: [[0,0], [0,0]]

If our output mask is as follows: [[255,0], [0,255]]

To represent this mask we have to first flatten the image into a 1-D array. this would result in
something like [255,0,0,255] for mask. then we can use the index to create the mask. finally we
would have something like [1,0,0,1] as our mask.

How to download the dataset

We will be using Kaggle API to download the dataset

1. Install kaggle package in python using !pip install kaggle.You can check if its already present using !pip show kaggle.
2. Upload the kaggle.json file which contains the api key to root directory ~/.kaggle/kaggle.json
3. Use !kaggle datasets download -d ashkhagan/brain-tumor-dataset-with-saliency to download it in zip format

Loading/Analysing/Cleaning the data

1. The dataset is in zip folder so first we need to unzip the dataset which can simply be done by a. !unzip /content/brain-tumor-dataset-with-saliency.zip
2. First we need the natsort library which is used to perform natural sort on strings i.e 1,2,3,4,10 instead of 1,10,2,3,
3. Since correspondence of images and masks are messed up natsort will bring them in order
4. We need to eliminate folders from images and masks that can be done by running a loop and only including files and skipping folders a. os.path.isfile(img_path + '/' + i) if this is false we skip because path is a folder
5. Then we declare np array of all zeros which has the shape of (images,image_dimensions)
6. Then we run a loop loading and processing each image before storing it in X,Y which contain the brain images and masks respectively

Image preprocessing

1. Initially sizes of images is 512*512*3 which we change to 256*256*1 grayscale
2. We also normalize the pixel intensities by bringing them b/w 0 and 1 by dividing them by 255.
3. 255.0 because they need to be in float

What are Saliency Maps?

1. Saliency Map is an important concept of deep learning and Computer vision. While

training images of birds how does CNN knows to focus on bird-related pixels and ignore
the leaves and the other background things in the image? By using the concept of
Saliency Map.

2. Saliency Map is an image in which the brightness of a pixel represents how salient the

pixel is i.e brightness of a pixel is directly proportional to its saliency. It is generally a
grayscale image.

3. Saliency maps are also called as a heat map where hotness refers to those regions of the image which have a big impact on predicting the class which the object belongs to.

4. Here is an example, the picture shown in the right is the saliency map of the left one which shows the regions which are more attentive part to CNN.

UNET ARCHITECTURE

1. The model architecture is fairly simple:

a. an encoder (for downsampling)
b. a decoder (for upsampling) with skip connection

2. it shapes like the letter U hence the name U-Net.

resUNET ARCHITECTURE

Why are we using Dice score instead of binary cross entropy

1. The main reason that people try to use dice coefficient or IoU directly is that the actual goal

is maximization of those metrics, and cross-entropy is just a proxy which is easier to
maximize using backpropagation.

2. In addition, Dice coefficient performs better at class imbalanced problems by design:
3. More advanced reasons can be found here -

https://stats.stackexchange.com/questions/321460/dice-coefficient-loss-function-vs-cross
-entropy

4. Even on checking the training stats you will notice that accuracy has be pretty much constant since the start only reason but dice score has been constantly improving

Measures to prevent overfitting and improve score

1. I am using Early stopping on val_dice_loss to make sure that score doesn’t get bad

2. Score can be improved to become way better by using more epochs I just used 75 due to time constraints 300 is probably a good number

NOTE - the last graphs of both UNET and RESUNET are plotted between epochs and dice_score apologies for the typo

Training statistics visualized of UNET

Training statistics visualized of resUNET

COMPARING results

UNET MODEL		RES-UNET MODEL

Visualizing the results