ABSTRACT

In order to extract useful data for diagnosis and therapy planning, medical image

segmentation is essential. In the framework of unsupervised learning, this paper suggests a

unique method for medical image segmentation utilizing the Gaussian Mixture Model. The

GMM is a probabilistic model that can capture intricate patterns and variations found in

medical pictures since it reflects a blend of Gaussian distributions.

Medical Image Segmentation is used to extract regions of interest (ROI) from medical

imagery which can be helpful for doctors to study the anatomy of the segmented regions to

find tumors etc. This paper uses the variational level sets method to perform image

segmentation which tries to find the best boundaries between different parts of an image by

minimizing the function which measures how good the segmentation is. The mentioned

function has two terms the first term deals with the edges of the image and the second term

deals with regions of the image and tries to make the regions have similar properties such as

color or texture. Optimizing the model to fit the observed pixel intensities in the medical

pictures, the GMM parameters are estimated using the Expectation-Maximization approach.

This paper is based on a statistical model called the Gaussian Mixture Model and also it

includes a new region-based term that uses GMMs of different regions to measure how

different they are from each other. Distance between two GMMs is calculated and this distance

is maximized during the segmentation process which makes the regions more distinct and

easier to separate. The proposed method tests the implementation of segmentation on

dermoscopy images and shows that its method can segment these images more efficiently.

1
 CHAPTER 1

INTRODUCTION

1.1 MEDICAL IMAGE SEGMENTATION:

Utilizing a probabilistic model to automatically partition an image into

distinct regions or structures without depending on previously labelled data is
known as medical image segmentation using the Gaussian Mixture Model
(GMM) in unsupervised learning. The underlying statistical characteristics of the
image are captured by the GMM, which operates on the assumption that the pixel
intensities in the picture adhere to a mixture of Gaussian distributions. When
getting labelled training data is difficult or unfeasible, this method is especially
helpful. Precisely identifying anatomical or diseased regions in medical pictures
is the ultimate goal of the GMM-based segmentation process, which comprises
phases including preprocessing, feature extraction, model training,
segmentation, and post-processing.

Gaussian Distribution: If all the data points in a dataset are symmetrically

distributed when plotted on a graph and the data points follow a bell shape then
the data points are said to be in Gaussian or normal distribution.

In normal or Gaussian distribution, Mean, Median, and Mode are the same
and Gaussian distribution can be described by two values: mean and standard
deviation. Smaller standard deviation results in a narrow bell curve and larger
standard deviation leads to a wide curve.

2
1.2 MACHINE LEARNING

GAUSSIAN MIXTURE MODEL:

Modelling the distribution of data in machine learning often involves the use of

probabilistic models such as Gaussian Mixture Models (GMMs). It is assumed that

many Gaussian distributions, each with parameters like mean, covariance, and weight,

combine to form the data. By learning these parameters during training, the model

becomes capable of representing multimodal or multimode complicated data

distributions. Applications for GMMs can be found in a number of domains, such as

speech recognition, image processing, and clustering, where they are used for tasks

including classification, density estimation, and data representation.

Problems with this method include figuring out how sensitive the medical images

are to initialization, noise, and intensity changes, as well as how many Gaussian

components is the right amount. To get meaningful and clinically relevant segmentation

findings, a successful implementation necessitates careful evaluation of these issues

and model parameter modification.

EXPECTATION-MAXIMIZATION FUNCTION:

In the framework of models such as the Gaussian Mixture Model (GMM), medical

picture segmentation frequently employs the statistical technique of expectation-

maximization (EM). To estimate parameters in probabilistic models, it functions as an

iterative method.

By improving its estimations iteratively, EM helps with medical picture

segmentation and facilitates model parameter improvement. Iteratively modifying the

3
model parameters to optimize the likelihood of the observed data entails assigning

probabilities to data points that correspond to various classifications. As EM helps the

algorithm better understand and adapt to complicated patterns in images, like those

present in medical imaging, it plays a critical role in improving the accuracy of

segmentation results.

1.3 OBJECTIVES:

• Achieve automatic and data-driven segmentation of medical images without

the need for pre-labeled training data.

• To divide the medical image into different components or clusters by image

segmentation using Gaussian Mixture Model.

• To identify specific regions of interest (ROI) within medical images for

detailed analysis.

• Utilize the probabilistic modeling capabilities of GMM to capture the

underlying statistical properties of pixel intensities in different regions of the

medical images.

• Enhance robustness to noise and variations in image intensity by modeling

the pixel intensity distribution with a mixture of Gaussian components.

• Develop an efficient and scalable segmentation approach suitable for

processing large volumes of medical images.

• To compare the segmentation results obtained using the GMM-based

approach with other state-of-the-art segmentation methods, quantitatively

measuring performance metrics such as Dice similarity coefficient,

sensitivity, specificity, and computational efficiency.

4
 CHAPTER 2

LITERATURE REVIEW

2.1 Existing System:

2.2 Drawbacks:

5
 CHAPTER 3
PROPOSED SYSTEM

• The proposed novel image segmentation method makes use of both edge-based

features and region-based features, unlike DRLSE or C-V which only make use of

edge-based or region-based respectively. The steps involved in the process are:

• Curve Initialization: Consider image I which consists of both object (Ro) and

background (Rb) and begins with a contour or curve around the object of interest.

• Curve Evolution: According to this model it is assumed that an image is modeled

as a mixture of Gaussian distributions of both object and background and we need to

modify the contour in such a way that maximizes the separation between object and

background using a metric called Bhattacharya distance which tells about the

dissimilarity between object and background distributions.

• Formulate an objective function that combines both edges-based and region-based

terms and optimize it during segmentation.

• Minimize the objective function by using differential calculus to find the optimal

configuration of contour that provides the best segmentation result and include the

novel term that maximizes the distance between object and background with level

sets to get a hybrid approach to image segmentation which considers both edge

details and regional characteristics in an image.

6
3.1 METHODOLOGY:

3.2 DATASET OVERVIEW:

7
3.3 IMPLEMENTATION MODULES:

Data Collection and Preprocessing:

• Standardize or normalize the medical image's intensity levels.

• Use techniques for enhancement or filtering to optionally improve the image quality.

Feature extraction from images:

• Extract relevant features from the image, such as pixel intensities or texture features.

Applying GMM:

• Apply the GMM algorithm to the extracted features.

• GMM estimates the parameters (mean, covariance, and weight) of the Gaussian
distributions that best fit the data.

8
Segmentation:

• Assign each pixel in the image to a particular cluster or class using the trained GMM.
• Using the acquired Gaussian distributions, this step efficiently divides the image into
various regions.

Evaluation:

• Utilize post-processing methods to improve the segmentation results, such as region

merging/splitting or morphological operations.
• Using relevant metrics, such as the Dice coefficient or Jaccard index, evaluate the
segmentation results' quality.

9
 CHAPTER 4

DESIGN

4.1 UML Diagrams:

Class Diagram:

10
Activity Diagram:

11
Sequence Diagram:

12
 CHAPTER 5

IMPLEMENTATION

5.1 ALGORITHM:

13
Terminologies used:

14
5.2 SOFTWARE AND HARDWARE REQUIREMENTS:

HARDWARE REQUIREMENTS:

Processor : Intel Core i5 or equivalent

RAM : 4GB or higher

Hard Disk Space : At least 30GB of free disk space.

SOFTWARE REQUIREMENTS:

Operating System : Windows 7 or later, macOS 10.9 or later, or Linux (64-bit)

Technology : Python 3.x (preferably 3.6 or later) installed on your system

IDE :

Dataset :

Libraries : NumPy, Matplotlib, Scikit learn

TECHNOLOGY DESCRIPTION:

JUPYTER:

Jupyter Notebook (previously ipython Notebooks) is an electronic intelligent

computational climate for making Jupyter scratch pad reports. The "note pad" term can

conversationally make reference to various substances, mostly the Jupyter web

application, Jupyter Python web server, or Jupyter record design contingent upon setting.

As indicated by the authority site of Jupyter, Project Jupyter exists to foster open-source

programming, open-guidelines, and administrations for intelligent figuring across many

programming dialects. Jupyter Book is an open-source project for building books and

archives from computational material. It permits the client to develop the substance in a

15
combination of Markdown, a lengthy form of Markdown called MyST, Maths and

Equations utilizing MathJax,

Jupyter Notebooks, reStructuredText, the result of running Jupyter Notebooks at assemble

time.

PYTHON:

Python is an amateur agreeable backend programming language. Python is like Ruby in

numerous viewpoints, but it is somewhat less verbose than other codes. Python is easy to

learn. You can foster a helpful device in Python regardless of whether you haven't taken a

software engineering class. You don't need to manage the lower-level parts of

programming, like memory the board, since it's undeniable level. Python might be utilized

to compose scripts, scratch sites, and make informational indexes. Python is generally

utilized in established researchers for logical figuring, and libraries exist to empower

sharing scholastic code projects in Python straightforward. Python is a web programming

language, subsequently it utilizes the web to impart. It comprehends how to send and get

web demands. "Inexactly composed" is the manner by which Python is depicted. This

classification of programming dialects needn't bother with you to announce the sort of

significant worth a capacity returns or the kind of factor before you make it when you

characterize it. Python is intuitive as in you can sit at a Python brief and compose your

projects by collaborating straightforwardly with the mediator. Python is Object-Oriented

Python upholds the ObjectOriented programming style or approach, which exemplifies

code inside objects. Python is a Fantastic Language for Beginners Python is an

extraordinary language for amateurs since it permits you to make a wide scope of projects,

from basic text handling to internet browsers and games.

16
 CHAPTER 6

EVALUATION DETAILS

6.1 Sample code

import numpy as np
from utils import COLORS, load_image
from scipy.stats import multivariate_normal
from sklearn.cluster import KMeans
from matplotlib import pyplot as plt
class GMM:
def _init_(self, ncomp, initial_mus, initial_covs, initial_priors):
self.ncomp = ncomp
self.mus = np.asarray(initial_mus)
self.covs = np.asarray(initial_covs)
self.priors = np.asarray(initial_priors)
def inference(self, datas): # E-step
unnormalized_probs = []
for i in range(self.ncomp):
mu, cov, prior = self.mus[i, :], self.covs[i, :, :], self.priors[i]
unnormalized_prob = prior * multivariate_normal.pdf(datas, mean=mu,
cov=cov)
unnormalized_probs.append(np.expand_dims(unnormalized_prob, -1))
preds = np.concatenate(unnormalized_probs, axis=1)
log_likelihood = np.sum(preds, axis=1)
log_likelihood = np.sum(np.log(log_likelihood))
preds = preds / np.sum(preds, axis=1, keepdims=True)
return np.asarray(preds), log_likelihood
def update(self, datas, beliefs):

17
# M-step
new_mus, new_covs, new_priors = [], [], []
soft_counts = np.sum(beliefs, axis=0)
for i in range(self.ncomp):
new_mu = np.sum(np.expand_dims(beliefs[:, i], -1) * datas, axis=0)
new_mu /= soft_counts[i]
new_mus.append(new_mu)
data_shifted = np.subtract(datas, np.expand_dims(new_mu, 0))
new_cov = np.matmul(np.transpose(np.multiply(np.expand_dims(beliefs[:, i],-1),
data_shifted)), data_shifted)
new_cov /= soft_counts[i]
new_covs.append(new_cov)
new_priors.append(soft_counts[i] / np.sum(soft_counts))
self.mus = np.asarray(new_mus)
self.covs = np.asarray(new_covs)
self.priors = np.asarray(new_priors)
if _name_ == '_main_':
# Load image
image_name = raw_input('Input the image name: ')
image_path = 'images/{}.jpg'.format(image_name)
image = load_image(image_path)
image_height, image_width, image_channels = image.shape
image_pixels = np.reshape(image, (-1, image_channels))
_mean = np.mean(image_pixels,axis=0,keepdims=True)
_std = np.std(image_pixels,axis=0,keepdims=True)
image_pixels = (image_pixels - _mean) / _std
# Normalization
# Input number of classes
ncomp = int(input('Input number of classes: '))

18
# Apply K-Means to find the initial weights and covariance matrices for GMM
kmeans = KMeans(n_clusters=ncomp)
labels = kmeans.fit_predict(image_pixels)
initial_mus = kmeans.cluster_centers_
initial_priors, initial_covs = [], []
for i in range(ncomp):
datas = np.array([image_pixels[j, :] for j in range(len(labels)) if labels[j] == i]).T
initial_covs.append(np.cov(datas))
initial_priors.append(datas.shape[1] / float(len(labels)))
# Initialize a GMM
gmm = GMM(ncomp, initial_mus, initial_covs, initial_priors)
# EM Algorithm
prev_log_likelihood = None
for i in range(1000):
beliefs, log_likelihood = gmm.inference(image_pixels) # E-step
gmm.update(image_pixels, beliefs) # M-step
print('Iteration {}: Log Likelihood = {}'.format(i+1, log_likelihood))
if prev_log_likelihood != None and abs(log_likelihood - prev_log_likelihood) <1e-10:
break
prev_log_likelihood = log_likelihood
# Show Result
beliefs, log_likelihood = gmm.inference(image_pixels)
map_beliefs = np.reshape(beliefs, (image_height, image_width, ncomp))
segmented_map = np.zeros((image_height, image_width, 3))
for i in range(image_height):
for j in range(image_width):
hard_belief = np.argmax(map_beliefs[i, j, :])
segmented_map[i,j,:] = np.asarray(COLORS[hard_belief]) / 255.0
plt.imshow(segmented_map)
plt.show()
19
 •
import numpy as np
from PIL import Image
COLORS = [
(255, 0, 0), # red
(0, 255, 0), # green
(0, 0, 255), # blue
(255, 255, 0), # yellow
(255, 0, 255), # magenta
]
def load_image(infilename) :
img = Image.open( infilename )
img.load()
data = np.asarray( img, dtype="int32" )
return data

20
 CHAPTER 7

FUTURE WORK
The Gaussian Mixture Model (GMM) of unsupervised learning has the potential to

facilitate numerous advancements and applications in the field of medical image

segmentation. Ongoing research and development are needed to improve the accuracy

and precision of GMM-based medical image segmentation, perhaps by utilizing novel

algorithms, model architectures, or optimization strategies. Investigating hybrid models

that draw on the advantages of both approaches to achieve more precise and efficient

segmentation, particularly when working with intricate and sizable medical imaging

datasets, by fusing GMM and deep learning techniques. GMM-based segmentation

techniques are modified to accommodate a range of medical imaging modalities,

guaranteeing stable performance on a variety of picture kinds, including CT, MRI, and

more. Real-time medical image segmentation is made possible by the optimization of

models and algorithms, which facilitates speedier analysis and clinical decision-making.

On-device processing capabilities for portable and point-of-care devices are also being

developed.

Unsupervised learning using GMM is applied to rare diseases where labeled data is

hard to come by. When looking for patterns and anomalies that might point to these kinds

of conditions, this can be especially helpful. Integration of GMM-based models into

current healthcare procedures and systems is ensured by their customization to meet

particular clinical workflows and needs. Working together internationally to make it

easier to share benchmarking standards, techniques, and datasets related to medical

imaging. This could help segmentation models based on GMMs become more

dependable and broadly applicable.

21
 CHAPTER 8

CONCLUSION

In conclusion, there are a lot of promising opportunities for advancement in the field of

medical image segmentation using Gaussian Mixture Models (GMM) in unsupervised

learning. The potential uses of GMM-based segmentation in the medical industry are numerous

and significant, especially as technology develops further. It is anticipated that practitioners

and researchers will concentrate on enhancing the GMM models' accuracy and precision,

investigating novel algorithms, and combining cutting-edge optimization strategies. Using the

complementary advantages of both techniques for improved segmentation performance, the

integration of GMM with deep learning approaches is probably going to be an important area

of research. One important feature that can increase the utility of GMM-based segmentation in

a variety of clinical scenarios is its ability to adapt to a wide range of medical imaging

modalities, including rare diseases. It is projected that on-device applications and real-time

processing capabilities will become indispensable for speedier analysis and decision-making,

particularly in point-of-care settings. In order to overcome the difficulties associated with a

lack of labelled data, transfer learning techniques may be essential in helping to train models

on bigger and more varied datasets. Additionally, by placing a strong emphasis on

explainability and interpretability, these models should be easier for clinicians to use and more

reliable, which will facilitate their integration into clinical workflows. Global cooperation and

data exchange programs will probably hasten the creation of globally applicable standards and

procedures as well as the development and validation of GMM-based segmentation models. To

guarantee fair and equitable outcomes for diverse patient populations, it is also anticipated that

ethical considerations and bias mitigation strategies will receive more attention.

22
 CHAPTER 9

REFERENCES

• Riaz, Farhan, et al. “Gaussian Mixture Model Based Probabilistic Modeling of

Images for Medical Image Segmentation.” IEEE Access, vol. 8, Institute of Electrical

and Electronics Engineers, Jan. 2020, pp. 16846–56.

https://doi.org/10.1109/access.2020.2967676.

• Raza, Khalid, and Nripendra Kumar Singh. “A Tour of Unsupervised Deep Learning

for Medical Image Analysis.” Current Medical Imaging Reviews, vol. 17, no. 9,

Bentham Science Publishers, Sept. 2021, pp. 1059–77.

https://doi.org/10.2174/1573405617666210127154257.

• Farnoush, R., and B. Zarr Pak. “Image Segmentation Using Gaussian Mixture

Model.” International Journal of Industrial Engineering & Production Research, vol.

19, no. 12, Mar. 2008, pp.29–32.

• www.researchgate.net/profile/Behnam_Zarpak/publication/250791017_Image_Seg

mentation_using_Gaussian_Mixture_Models/links/5580f85708ae47061e5f3f48.pdf

23