1 of 15

Brain Tumor Detection:

MRI Image Segmentation Using U-net

By: Peike Li, Queenie Wu, Juslyn Theriault, Eva Wade, and Aidan Perez

2 of 15

Overview

Background

Model Architecture

Model Building &�Training

Loss function and metrics

MRI

​

U-net

​

​

​

Model

​

Model Assessment

​

​

​

Learning curve &

conclusion�

Result

​

3 of 15

MRI & FLAIR

Abnormal watery region

  • MRI is a medical imaging technique used in radiology to form pictures of brain tissues

​

  • FLAIR is a type of MRI scans that detects the presence of fluid where it shouldn’t be

4 of 15

Our dataset

  • Our dataset contains 110 patients’ MRI scans obtained from The Cancer Imaging Archive
  • FLAIR + Mask
  • Masks are used as labels that highlight the area of interest (abnormal regions) in FLAIR
  • Mask is our ground truth “y variable”

​

5 of 15

Goal and Potential Applications

  • Goal: Automate the process of detecting abnormal areas in MRI using ML
  • Quantify the abnormal areas
    • Provide metrics like volume and spatial distribution to monitor disease progression
  • Enhance efficiency in reading scans
    • Highlight easily to reduce workload of health professionals

​

​

6 of 15

Solution

​

Image segmentation using convolutional neural network (CNN):

U-Net

7 of 15

U-Net Background

Original Application:

Segmenting Images of cells

​

​

Modern Application:

Generative AI

​

​

Pug Dancing

Ballet

​

​

8 of 15

U-Net

Encoder

Decoder

Skip Connection

Skip Connection

Skip Connection

Channels/Dimensions

9 of 15

Procedures

Procedure overview

Read Images

Image Augmentation

“An image to (256, 256, 3)”

–– OpenCV

“Rotate, zoom, shift, flip

Recognize me in all angles”

–– more training data

Train-Test

Split

Model

Building

Model

Training

& Assessment

“Encoder, Decoder, Me!”

–– U-Net

10 of 15

Model Building

Build a U-Net

Get Layer (256 x 256)

Get Layer (128 x 128)

Get Layer (64 x 64)

Get Layer (32 x 32)

Bridge

Pretrained Model As Encoder

Customized Decoder

Get Layer (32 x 32)

Get Layer (64 x 64)

Get Layer (128 x 128)

Get Layer (256 x 256)

11 of 15

Model Building

Build a U-Net

Get Layer (256 x 256)

Get Layer (128 x 128)

E

Get Layer (32 x 32)

Bridge

Pretrained Model As Encoder

Customized Decoder

D

Get Layer (64 x 64)

Get Layer (128 x 128)

Get Layer (256 x 256)

What’s happening in a decoder layer?

12 of 15

Inside a Decoder Layer…

D = Last decoder layer

Concatenate

E = encoder layer

E after integrating D

D after Upsampling

Input 1

Input 2

“Attention gate”

Convolutional Block

D_new = Decoder layer

Feature extraction + learning

13 of 15

Loss Function & Metrics

  • Dice Coefficient Loss Function
    • Determined by the Dice Coefficient metric (1 - Dice Coefficient)
  • Binary Cross-Entropy Loss function
  • Jaccard’s Index (IoU)

Dice Coefficient

Jaccard’s Index

14 of 15

Suppose model was trained successfully…

For illustration purpose

Learning Curve

Epochs

Truth-Prediction Comparison

15 of 15

Reference

https://www.kaggle.com/code/tajvarhirad/brain-image-segmentation/notebook

https://www.kaggle.com/code/promithaldar/attention-res-unet/notebook

​