You are on page 1of 5

Artificial intelligence (AI) holds the promise of revolutionizing healthcare, much as it has transformed various pattern

recognition applications. Over the past few years, there has been a growing focus on medical image analysis, with significant
strides made in machine learning (ML) research applied to medical images. The primary objective of this study was to assess
the classification accuracy of ML algorithms in clinical images.Utilizing three distinct feature extraction techniques—Local
Binary Patterns (LBP), Histograms of Gradients (HOG), and pixel-based extraction—our study compares seven classification
models. These models encompass Support Vector Machines (SVM), Decision Trees (DT), Logistic Regression (LR), Random
Forests (RF), Extreme Gradient Boosting (XGB), K-Neighbors Classifiers (KN), and Multinomial Naive Bayes (NB). The focus is
on classifying essential features obtained through diverse feature extractors.For a rigorous evaluation of our classification and
feature extraction models, we chose to test their accuracy using a publicly available dataset of COVID-19 chest radiographs.
This dataset comprises 21,212 CT images categorized into four classes, specifically tailored for histopathology images.Among
the various feature extractors examined, SVM yielded superior results when coupled with HOG features.

The main contributions of this study are as follows:

1. The comparison of different Machine Learning Algorithm for Medical images Using Pixel, LBP and HOG Feature Extraction.

2. A combination of HOG and SVM was proposed to classify three types of X-ray images: normal, COVID-19, Viral_Pnuemonia,
and lung_opacity. In comparison to other deep learning and machine learning models in the literature, our model achieves
competitive performance.

The significance of biomedical image analysis and processing in the medical field, particularly in
non-invasive treatment and clinical studies, cannot be overstated. Machine learning plays a
crucial role in healthcare informatics, offering applications in disease diagnosis, classification,
and prognosis. In the current era, machine vision has emerged as a vital tool in disease detection
and treatment, aligning seamlessly with advancements in medical science and technology. By
synergizing the efforts of medical professionals, surgeons, and scientific technology, we have the
potential to enhance efficiency, accuracy, and early disease detection. This collaborative
approach not only reduces the time involved but also contributes to minimizing the overall cost
of treatment.

3.1 Study Of Machine Learning (ML) Algorithms

3.3.1 Logistic Regression(LR): Statistical algorithms used for binary classification such as logistic regression predict whether
an event will occur according to [25] . This algorithm extracts a number of low-level features directly from raw, noisy pixels
and the edge detection algorithm is the most prevalent. The relationship between the properties of the input and the
expected output variables can be represented using the logistic function, also called the sigmoid function. The sigmoid
function ( ) is written as follows:

( ) (7)

where, (8)

For logistic regression, the likelihood function is as follows:


( ) ∏ ( ) ( ( )) (9)

3.3.2 Decision Tree(DT) : An algorithm called the DT was developed in 1984 [26]It is a ML and data mining algorithm [27]. Throughout
the DT, checks and compares the classifier with similarity in the dataset and ranks it according to its classification.[28] (Hussain et al.,
2020) used DTs to classify data based on selecting an attribute which optimizes and fixes the data division. Dataset[29] attributes are
divided into several classes until conclusion criteria and conditions are met. A mathematical representation of the DT algorithm is as
follows::
∑ (10)
3.3.3 Random Forest(RF): [30]developed RF as an ensemble learning(EL) method for solving classification and regression problems
Using multiple models to solve the same problem is ensemble learning, a machine learning method for boosting accuracy. A single
classifier can only provide a small amount of accuracy, whereas there are multiple classifiers participating in ensemble classification[31].
∑ ( ) (12)
In an uncertain situation,

(( )
√∑ )
(13)

3.3.4 Extreme gradient boosting(XGB): A distributed gradient boosting algorithm, XGBoost[32], optimizes gradient boosting for large
datasets. In general, the XGBoost[33] algorithm functions as follows:

Assume a training dataset

* + , (14)

then

* + (15)

where, m denotes to m dimension and,

∑ ( ) , (16)

In the model, represents the trees number.

3.3.5 K- Neighbors Classifier(KNN) : According to [34]the K-Nearest Neighbor classifier divides a dataset into multiple clusters
according to the user's preferences. As a result of its flexibility, this algorithm can be useful to classification as well as regression problems.
A KNN algorithm finds clusters by comparing the separation between different data points, based on the premise that similar items tend to
cluster together [35].

( ) √( ) ( ) (17)
In contrast to algorithms that learn directly from training datasets, KNN holds data and operates on it during the classification phase. It is
determined by the majority vote of its nearest neighbours whether a new data point should be classified. When k is changed, the algorithm's
accuracy can be affected[36].

3.3.6 Multinomial Naive Bayes(NB): As a probabilistic method that is independent of training and testing, NBM relies on a
probabilistic approach [34]. In order to classify patients according to their response to treatment, we used a multinomial
naive Bayes (MNB) classifier. This classifier makes use of a multinomial distribution to classify individual features. . For the
training process, suppose represents the positive and non-positive classes and is defined
as positive conditioning[37] factors ( ). This formula can be used to calculate the probability( ) of each event within a class
( | ) ∏ ( ( | )) (18)

( | )
( | ) (19)
By using the maximization function, a concept for indexing a query or a document is selected
( ) ( ) ∏ ( ( | )) (20)

3.3.7 Support Vector machine(SVM): In orthodox ML applications, a SVM can solve classification glitches more effectively than a
regular machine learning algorithm [38]. This method provides a nonlinear efficient way to approximate a problem. A support vector
machine creates a spreadable hyperplane by plotting the input data to a high-dimensional feature space and exploiting the margin between
two groups in that space. Maximizing a margin between two groups can be achieved using a quadratic system designed to solve Lagrangian
multipliers [39].
( ) (21)
Where, a = dimensional coefficient, c = offset
With SVM, the optimal hyperplane is shown by the dot product function of the high-dimensional feature space called kernels. In SVM,
which are approximately input points, can be used visualize the optimal hyperplane elucidation.

1 RESULTS AND EXPERIMENT

Hyperparameter Value
Activation Softmax, relu
Momentum 0.99
Epsilon 0.001
Optimizer Adamax
Epoch 20
Batch_size 64
Learning_rate 0.0001

LBP

1. Working Principle:
o In the context of chest X-ray images, the LBP approach involves analyzing the
local texture patterns within the lung regions.
o For each pixel in the X-ray image, the intensities of neighboring pixels within a
defined radius are compared to the central pixel.
o A binary pattern is formed based on whether the intensity of each neighbor is
greater than or equal to the central pixel's intensity.
o These binary patterns are then encoded and used to represent the local texture
information in the chest X-ray image.
2. Operation Steps:
o Define a neighborhood around each pixel in the chest X-ray image.
o Compare the intensity of each neighbor to the intensity of the central pixel.
o Generate a binary pattern based on the comparison results.
o Convert the binary pattern to decimal to create a numerical representation of the
local texture feature for each pixel.
o Apply LBP across the entire chest X-ray image to capture variations in texture
patterns.
3. Applications:
o LBP is utilized in chest X-ray images for tasks such as identifying specific texture
patterns related to abnormalities, nodules, or other pulmonary conditions.
o It aids in distinguishing between different tissue textures, providing valuable
information for disease diagnosis.

Histograms of Gradients (HOG) Approach for Chest X-ray Images:

1. Working Principle:
o When applied to chest X-ray images, the HOG approach involves capturing the
distribution of gradients, which represent changes in pixel intensities.
o The gradients are calculated for localized regions within the lung area of the X-
ray image.
o Histograms of these gradient orientations are constructed, providing a
representation of the structural features and patterns present in the lungs.
2. Operation Steps:
o Calculate the gradient magnitudes and orientations for each pixel in localized
regions of the chest X-ray image.
o Construct histograms based on the distribution of gradient orientations.
o Normalize the histograms to account for variations in illumination and enhance
the robustness of the approach.
3. Feature Representation:
o The resulting HOG feature vector encapsulates information about the shape and
structure of the lung regions in the chest X-ray image.
o It is particularly useful for capturing structural patterns indicative of different
pulmonary conditions.
4. Applications:
o HOG is employed in chest X-ray images for tasks such as detecting specific
structural features, identifying anomalies, or assisting in the diagnosis of
respiratory diseases.

Pixel Approach for Chest X-ray Images:

1. Working Principle:
o In the pixel approach for chest X-ray images, each pixel's intensity value is
directly used for analysis without additional transformation.
o The focus is on considering the individual pixel values to understand the overall
intensity distribution in the X-ray image.
2. Operation Steps:
o Treat each pixel in the chest X-ray image as a separate feature, considering its
intensity or grayscale value.
o Analyze the overall distribution of pixel intensities to discern patterns and
variations.
3. Feature Representation:
o Features are represented by the raw pixel values, and each pixel contributes
independently to the analysis.
4. Applications:
o The pixel approach is suitable for basic tasks in chest X-ray images, such as
identifying overall brightness, contrast, or basic structural features.
o It provides a straightforward analysis of the raw intensity values without intricate
transformations.

You might also like