Scribd
Upload
55 views

3d Image Segmentation

Uploaded by

abhishek dhangar
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
55 views

3d Image Segmentation

Uploaded by

abhishek dhangar
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 11

3D Image Segmentation using

UNETR Transformer

Akshay Dhangar, (2008390100004)

Rajkiya Engineering College, Kannauj


Major Project (KCS-753)
under the guidance of

Mr. Shashank Yadav


(Asisstant Professor)
Table Of Contents

Introduction
Literature Survey
Proposed Approach
Implementation
Result
Refrences
Introduction

Problem Statement
3D Image Segmentation using UNETR Transformer

Motivation
To aid in 3D medical scans like MRIs and CTs for better diagnosis
Introduction

3D image segmentation is the process of partitioning a 3D image into


multiple segments, or regions of interest, based on certain criteria.
FCN networks have been used predominantly in medical image
segmentation for long. Despite this,locality of convolutional layers
limits their learning capacity of long- range spatial dependencies.
UNETR uses transformer as encoder to learn sequence representations
of the input volume and capture the global multi-scale information,
“U-shaped” network design for the encoder and decoder.
Transformer encoder is directly connected to a decoder via skip
connections at different resolutions to compute the final semantic
segmentation output.
Literature Survey

Table: A Brief Survey I: Title, Models, Key Contribution


Title Models Key Contribution
U-Net: Convolutional UNET Archi- use in nearly all ma-
Networks for Biomed- ture jor image modalities,
ical Image Segmenta- from CT scans and
tion [1], 2015 MRI to X-rays and mi-
croscopy
Vision Transformers Vision good for large
for Lung Segmenta- Trannsformer datasets and uses
tion on CXR Images skip connections
[2], 2017
Image Segmentation ResNET 3D deep learning method
Using Deep Learning: abd uses resiudal con-
A Survey [3],2020 nection
Dataset and Parameters

DataSet is taken from Kaggle:


While original Clinical, genetic, and pathological data resides in the
Genomic Data Commons (GDC) Data Portal while the radiological data is
stored on The Cancer Imaging Archive (TCIA).

https://www.kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation.

Parameters:
Number of layers and neurons in a neural network
Learning rate
Batch size
Loss function
Threshold values, Neighborhood size
Image dimensions, Intensity and modality (e.g., CT, MRI)
Propsed Approach

We are using EDA analysis to train our Transformer Architecture


to:
Get maximum insights from a data set.
Uncover underlying structure.
Extract important variables from the dataset.
Detect outliers and anomalies.
Test underlying assumptions.
Determine the optimal factor settings.
Propsed Approach

Loss function
Dice Score/Coefficient The Dice coefficient is a popular segmentation
metric which measures the overlap between two samples. The metric
generates a score ranging between 0 and 1 inclusively where a Dice score
of 1 indicating perfect overlap and a score of 0 indicating no overlap. The
Dice score is obtained by computing the ratio of the intersection) of the
two samples and sum of their element-wise sum.
T
|ytarget ypred |
DiceScore = 2
|ytarget | + |ypred |
Result

Here I get loss is about 0.23, which is a better score in comparison to oter
transformer models without EDA analysis.
Some of other results are given:
Conclusion and future Scope

UNETR has demonstrated exceptional performance in 3D medical


image segmentation, achieving state-of-the-art results on various
datasets.
It effectively combines the strengths of U-Net’s encoder-decoder
architecture with Transformers’ ability to capture long-range
dependencies, leading to improved accuracy and robustness.
The future of UNETR and transformer-based models for 3D image
segmentation holds exciting possibilities for enhancing medical
diagnosis, treatment planning, surgical guidance, and advancements
in healthcare and life sciences.
Refrences

N. Siddique, S. Paheding, C. P. Elkin, and V. Devabhaktuni, “U-net


and its variants for medical image segmentation: A review of theory
and applications,” IEEE Access, vol. 9, pp. 82031–82057, 2021.
A. Hatamizadeh, Y. Tang, V. Nath, D. Yang, A. Myronenko,
B. Landman, H. Roth, and D. Xu, “Unetr: Transformers for 3d
medical image segmentation,” 2021.
J. Zhang, F. Bai, J. Zhao, and Z. Song, “Multi-views action
recognition on 3d resnet-lstm framework,” in 2021 IEEE 2nd
International Conference on Big Data, Artificial Intelligence and
Internet of Things Engineering (ICBAIE), pp. 289–293, 2021.

You might also like