Experiment 14

Implementation of Machine Learning Classifiers on Medical

Imaging Dataset using Python.
Abdul Rehman Ahsan Naveed Arslan Siddique Talha Imad
Department of Biomedical Engineering Department of Biomedical Engineering Department of Biomedical Engineering Department of Biomedical Engineering
Air University Air University Air University Air University
Islamabad,Pakistan Islamabad,Pakistan Islamabad,Pakistan Islamabad,Pakistan
[email protected] [email protected] [email protected] [email protected]

1. Traditional Machine Learning Classifiers

Abstract - The implementation of machine learning
classifiers on medical imaging datasets using Python has  Support Vector Machines (SVM):
emerged as a transformative approach in healthcare SVMs are widely used for binary classification tasks
diagnostics, enabling automated, accurate, and efficient in medical imaging, such as tumor detection. They
analysis of complex medical images. By leveraging powerful perform exceptionally well in high-dimensional
Python libraries such as TensorFlow, Scikit-learn, and spaces and can handle small datasets effectively.
PyTorch, classifiers like Support Vector Machines, Random SVMs rely on kernels (e.g., radial basis function
Forests, and Convolutional Neural Networks are trained on kernels) to map input data into higher dimensions for
datasets to identify patterns and anomalies, including tumor better separation of classes.
detection, disease classification, and segmentation tasks.  Random Forest (RF):
This study underscores the critical role of preprocessing Random Forest is an ensemble learning method that
techniques, such as normalization and data augmentation, creates multiple decision trees and aggregates their
to enhance model performance, while emphasizing the predictions. It is particularly useful for feature
importance of evaluation metrics like accuracy, precision, selection and classification in medical imaging, such
and recall to ensure reliability. The fusion of Python's as detecting diabetic retinopathy or identifying breast
computational capabilities with advanced machine learning cancer lesions.
techniques facilitates the development of scalable,  K-Nearest Neighbors (KNN):
reproducible, and clinically relevant solutions, paving the KNN is a simple yet effective classifier that assigns a
way for improved diagnostic accuracy and personalized class based on the majority vote of its nearest
medicine. neighbors. It has been applied in segmenting medical
images and classifying diseases in relatively small
datasets.
I. INTRODUCTION
 Naive Bayes:
Naive Bayes classifiers are probabilistic models often
A. Background employed for textural analysis in medical images,
such as predicting disease presence from MRI scans
Medical image processing is a cornerstone of modern based on pixel intensity distributions.
diagnostics, playing a pivotal role in detecting, classifying,
and monitoring diseases. With the ever-growing complexity 2. Deep Learning Classifiers
and volume of medical imaging data generated by modalities
like MRI, CT, X-ray, and ultrasound, traditional manual  Convolutional Neural Networks (CNN):
analysis methods are increasingly inadequate. Human analysis, CNNs are the backbone of modern medical image
while effective, is often time-consuming, subjective, and processing. Their ability to learn spatial hierarchies
prone to inter-observer variability. This creates a pressing of features makes them ideal for tasks like
need for automated, robust, and scalable solutions. Machine segmentation (e.g., U-Net), classification, and
learning (ML) classifiers have emerged as transformative anomaly detection. CNNs have demonstrated state-
tools, leveraging computational power to enhance the speed, of-the-art performance in detecting lung nodules,
accuracy, and efficiency of medical image interpretation. brain tumors, and retinal diseases.
 Recurrent Neural Networks (RNN):
Machine learning classifiers function as decision-making While less common in medical imaging, RNNs are
models that learn patterns from training data and generalize used in tasks involving temporal data or sequences,
their knowledge to unseen data. Their ability to handle large, such as dynamic imaging or real-time image analysis.
high-dimensional datasets makes them well-suited for medical  Deep Belief Networks (DBN):
imaging applications. Various classifiers, ranging from These unsupervised deep learning models are useful
traditional algorithms to modern deep learning models, for feature extraction and dimensionality reduction in
contribute to this field: complex medical imaging datasets.
 Generative Adversarial Networks (GAN):
GANs play a significant role in medical imaging for
 generating synthetic data, enhancing image planning or volumetric analysis (e.g., CNNs,
resolution, and augmenting datasets to improve RF).
classifier performance.  Vascular Structures:
o Segmentation of blood vessels in
3. Hybrid and Ensemble Approaches angiograms or retinal images for diagnosing
vascular diseases (e.g., CNNs, GANs).
Hybrid models combine traditional machine learning
techniques with deep learning for improved performance. For 3. Anomaly Detection
example, CNNs can be used for feature extraction, followed
by an SVM or Random Forest classifier for final decision-  Fractures and Bone Abnormalities:
making. Ensemble methods, like stacking and boosting, o Detection of fractures in X-rays (e.g., SVMs,
aggregate multiple classifiers to reduce bias and variance, CNNs).
achieving more reliable results in challenging tasks such as  Pulmonary Diseases:
multi-class disease classification. o Identification of pneumonia, tuberculosis,
and COVID-19-related lung abnormalities
using chest X-rays and CT scans (e.g.,
Random Forest, CNNs).
B. Problem Identification
 Neurological Disorders:
o Detection of stroke, multiple sclerosis
In medical image processing, using various machine learning lesions, and brain hemorrhages from MRI
classifiers allows for the identification of a wide range of scans (e.g., SVMs, CNNs).
diseases, conditions, and anatomical features. Below is a
detailed breakdown of what these classifiers can identify,
grouped by medical imaging applications: 4. Functional Analysis

1. Disease Detection and Classification  Cardiac Function:

o Quantification of ejection fraction and
 Cancer Diagnosis: ventricular wall motion abnormalities in
o Lung Cancer: Detection of nodules and echocardiograms (e.g., CNNs, RF).
classification of malignancy using CT scans  Lung Function:
(e.g., CNNs). o Analysis of lung volumes and air trapping in
o Breast Cancer: Identification of masses, CT scans for conditions like COPD (e.g.,
microcalcifications, and classification into CNNs, SVMs).
benign or malignant using mammography
(e.g., SVM, Random Forest). 5. Prognostic Analysis
o Brain Tumors: Differentiation of gliomas,
meningiomas, and metastases using MRI  Disease Progression:
scans (e.g., CNNs, RF). o Predicting tumor growth rates and
 Cardiovascular Diseases: metastasis from sequential imaging (e.g.,
o Detection of coronary artery blockages, RNNs, Time-series CNNs).
arrhythmias, and myocarditis from imaging  Survival Predictions:
modalities like echocardiography and MRI o Analysis of imaging biomarkers to predict
(e.g., CNNs, SVMs). survival outcomes in cancer patients (e.g.,
 Diabetic Retinopathy: SVM, RF).
o Identification of retinal hemorrhages and
microaneurysms from fundus images (e.g., 6. Anatomical Feature Recognition
Random Forest, CNNs).
o Skin Cancer:
o Classification of melanomas and other skin  Landmark Detection:
lesions using dermoscopic images (e.g., o Identification of anatomical landmarks in
CNNs). skeletal images for orthodontics, prosthetics,
and surgical navigation (e.g., KNN, SVMs).
 Alzheimer’s Disease:
o Detection of cortical thinning, hippocampal  Age Estimation:
atrophy, and amyloid plaque deposits in o Bone age prediction from hand X-rays for
MRI and PET scans (e.g., SVM, DBNs). pediatric assessments (e.g., CNNs).

2. Image Segmentation 7. Workflow Optimization

 Tumor Segmentation:  Quality Control:

o Precise delineation of tumors in organs like o Detection of poor-quality scans or artifacts
the brain, liver, and lungs to aid in treatment to ensure optimal image analysis (e.g.,
planning (e.g., U-Net, Mask R-CNN). SVMs, RF).
 Organ Segmentation:  Scan Triage:
o Identifying boundaries of organs such as the o Prioritization of critical cases, such as
heart, liver, and kidneys for surgical identifying urgent abnormalities like brain
hemorrhages (e.g., CNNs, GANs).
8. Rare and Genetic Disorders 5. Generative AI for Investigating Medical Imagery
Models: Researchers presented a framework leveraging
 Congenital Abnormalities: generative AI to understand AI models in medical
o Detection of structural anomalies like cleft imaging. This approach helps identify and interpret
palate or congenital heart defects from visual cues associated with model predictions, enhancing
ultrasound or MRI (e.g., CNNs). transparency and trust in AI-driven diagnostic tools.
 Genetic Disease Markers: 6. Federated Learning Combined with Transfer
o Analysis of characteristic imaging patterns Learning: An integrated approach combined federated
associated with conditions like Marfan learning with transfer learning for brain MRI
syndrome or Duchenne muscular dystrophy classification. This method enables decentralized model
(e.g., CNNs, SVMs). training across multiple clients without compromising
data privacy, addressing critical confidentiality concerns
in medical data handling.
9. Image Enhancement and Reconstruction
7. Self-Supervised Learning in Medical Image
Classification: A study investigated self-supervised
 Noise Reduction and Super-resolution: learning techniques for medical image classification,
o GANs and autoencoders improve image achieving state-of-the-art performance with
quality by reducing noise and enhancing approximately 100 labeled training samples per class.
resolution for better diagnosis. This approach significantly reduces the dependency on
 Missing Data Recovery: large labeled datasets.
o Reconstruction of incomplete images, such 8. Tensor Networks for Medical Image Classification:
as partially scanned MRI data, for Research explored the use of tensor networks,
diagnostic utility. traditionally utilized in quantum physics, for medical
image classification. The proposed locally orderless
C. Other Researches tensor network model (LoTeNet) demonstrated
performance comparable to state-of-the-art deep learning
Recent research has extensively explored the application of methods, with fewer hyperparameters and reduced
machine learning classifiers in medical image processing, computational resources.
leading to significant advancements in diagnostic accuracy 9. Machine Learning Classification in Mammography
and efficiency. Notable studies include: Using BI-RADS: A study evaluated the classification
accuracy of various state-of-the-art image classification
models across different categories of breast ultrasound
images, as defined by the Breast Imaging Reporting and
Data System (BI-RADS). The findings indicated the
1. Semi-Supervised Learning for Pathological Image
potential for enhanced diagnostic accuracy in breast
Classification: A study proposed a semi-supervised
imaging.
learning method that utilizes a small amount of labeled
pathological image data to train a network model. The
model integrates features extracted by the network to These studies collectively underscore the transformative
classify images, demonstrating effectiveness in scenarios impact of machine learning classifiers in medical image
with limited labeled data. processing, paving the way for more accurate, efficient, and
2. Shallow vs. Deep Learning Classifiers in Medical personalized diagnostic tools.
Imaging: Research compared traditional (shallow)
machine learning classifiers with deep learning models D. Our Approach
in medical image analysis. The study highlighted that
while deep learning models, such as Convolutional
Neural Networks (CNNs), automatically extract and a) Understanding Data Preprocessing
classify features, shallow classifiers require manual
feature extraction. The importance of model  The importance of preprocessing techniques such as
explainability was emphasized for integration into normalization, noise reduction, and augmentation to
clinical practice. improve model performance and robustness.
3. Uncertainty Quantification in Classifier Predictions:
A study introduced an uncertainty metric for machine
learning classifiers applied to medical imaging datasets. b) Feature Engineering and Representation
This metric assesses the reliability of classifier
predictions on individual images, aiding in clinical  How features are extracted, selected, or learned
decision-making by indicating the confidence level of automatically from imaging data to improve
each prediction. classification accuracy and efficiency.
4. Deep Learning Techniques for Medical Image
Classification: A comprehensive review examined c) Role of Different Classifiers
various deep learning methods for medical image
classification, focusing on their performance across  Gaining insights into the strengths and limitations of
different imaging modalities. The review concluded that various classifiers:
deep learning classifiers are highly effective in
addressing classification challenges in medical image
processing. o Traditional models like SVMs for smaller
datasets.
 o Deep learning models like CNNs for Step 2: Load the Dataset
complex, high-dimensional image data.
o Hybrid models for specific challenges in Here, a preloaded medical dataset, such as the breast cancer
medical imaging. dataset, is used. This dataset contains features (numerical
attributes derived from medical imaging or diagnostics) and
d) Evaluation Metrics and Model Validation corresponding target labels (e.g., benign or malignant tumors).
This dataset is split into two parts:
 Learning the significance of metrics such as accuracy,
precision, recall, F1 score, ROC-AUC, and confusion 1. Features (X): The numerical attributes or predictors.
matrices in evaluating classifier performance. 2. Target (y): The classification labels or outcomes.
e) Challenges in Real-World Applications Step 3: Split Dataset into Training and Testing Sets

 Understanding practical issues such as class The dataset is divided into two subsets:
imbalance, lack of labeled data, and ethical
considerations (e.g., data privacy and bias).
 Training Set: Used to train the machine learning
classifiers.
f) Impact on Personalized Medicine  Testing Set: Used to evaluate the model's
performance. Typically, 80% of the data is used for
 Machine learning’s role in tailoring diagnostics and training, and 20% is used for testing. The split
treatment plans to individual patients based on ensures that the model is evaluated on data it has
imaging biomarkers and classifier predictions. never seen before, simulating real-world scenarios.

g) Ethical and Regulatory Considerations Step 4: Standardize the Features

 The importance of ensuring model explainability, Medical data often involves attributes on different scales (e.g.,
transparency, and compliance with medical image intensity values vs. feature size). Standardization
regulations to foster trust and adoption in clinical ensures that all features have a mean of 0 and a standard
practice. deviation of 1. This step is crucial for algorithms like SVM
and k-NN that are sensitive to feature scaling.
II. MATERIAL & METHOD
Step 5: Train the k-Nearest Neighbors (k-NN) Classifier

A. Material:
The k-NN algorithm is a simple yet effective method where:

Google Colab is a cloud-based platform that provides a  The classifier identifies the 'k' nearest data points in
Jupyter notebook environment for writing and executing the training set to a new test point.
Python code. It is popular for data analysis, machine learning,  It assigns the most common label among the
and medical image processing due to its free access to neighbors to the test point.
computational resources like GPUs and TPUs. Colab
simplifies the process of working on Python projects,
especially when dealing with large datasets or The model is trained on the standardized training data and
computationally intensive tasks. Its integration with Google then used to make predictions on the test set. The choice of 'k'
Drive allows seamless storage and sharing of data, making it a (e.g., 5) determines the number of neighbors considered.
valuable tool for collaborative research and development.
Step 6: Train the Support Vector Machine (SVM)
B. METHOD: Classifier

Step 1: Import Libraries SVM is a more sophisticated algorithm that works by finding
the optimal hyperplane to separate classes in the feature space:
This step involves bringing in essential Python libraries that
enable various functionalities:  With a linear kernel, the model assumes that the
classes can be separated by a straight line or
hyperplane.
 Numpy: Handles numerical operations and data  The algorithm finds the hyperplane that maximizes
arrays. the margin (distance) between classes, making it
 Matplotlib: Used for visualizations, such as plotting robust to outliers.
results.
 Scikit-learn: Provides tools for dataset loading,
The SVM model is trained and tested similarly to the k-NN
preprocessing, model building, and evaluation.
model.
Step 7: Train the Random Forest (RF) Classifier o Accuracy: Moderate to high, depending on
the dataset structure.
Random Forest is an ensemble learning method: o Confusion Matrix: Indicates how well the
classifier distinguished between benign and
malignant cases.
 It constructs multiple decision trees during training
o Classification Report: Precision and recall
and combines their predictions (majority voting for
may vary, especially if the classes are
classification) to improve accuracy and prevent
imbalanced.
overfitting.
 Each tree is trained on a random subset of data,
making the model robust and less prone to bias from
individual features.

The RF model is trained on the standardized data and tested

on the test set.

Step 8: Make Predictions

For each classifier (k-NN, SVM, RF): [ Fig 1 knn metrics ]

1. Use the trained model to predict the labels of the test 2. Support Vector Machine (SVM)
set.
2. Compare the predicted labels with the actual labels in  Strengths: Finds an optimal hyperplane to maximize
the test set to evaluate performance. the margin between classes, ensuring robustness and
generalization.
Step 9: Evaluate Performance  Weaknesses: May struggle with very large datasets
or overlapping classes without kernel tuning.
The performance of each classifier is evaluated using:  Results:
o Accuracy: Generally high, especially with a
linear kernel for linearly separable data like
1. Accuracy: Measures the proportion of correct the breast cancer dataset.
predictions. o Confusion Matrix: Often exhibits minimal
2. Confusion Matrix: Provides insights into true false positives and false negatives,
positives, true negatives, false positives, and false indicating reliable classification.
negatives. o Classification Report: Precision and recall
3. Classification Report: Includes metrics like are typically balanced, showing good
precision, recall, and F1-score for each class. overall performance.

Step 10: Compare Results

Finally, the results of all three classifiers are compared:

 k-NN: May perform well on simpler datasets but can

struggle with noisy data or when feature scaling is
inadequate.
 SVM: Often achieves high accuracy with linear
separable data but may require more computational
resources.
 Random Forest: Typically provides robust and [ Fig 2 SVM metrics ]
accurate results, especially on complex or
imbalanced datasets, but may require tuning of 3. Random Forest (RF)
hyperparameters for optimal performance.
 Strengths: Robust against overfitting, especially for
III. RESULTS AND OBSERVATIONS complex or imbalanced datasets. It aggregates
multiple decision trees for reliable predictions.
1. k-Nearest Neighbors (k-NN)  Weaknesses: Computationally intensive with a large
number of trees and may require hyperparameter
 Strengths: Simple to understand and effective for tuning for best results.
smaller datasets. It relies on proximity, making it a  Results:
good baseline model. o Accuracy: Often the highest among the
 Weaknesses: Sensitive to feature scaling and high- three classifiers due to its ensemble nature.
dimensional data, prone to being affected by noisy or o Confusion Matrix: Shows strong
irrelevant features. performance in correctly identifying both
benign and malignant cases.
 Results:
 o Classification Report: Precision, recall, and security, and integrating multi-modal datasets to unlock the
F1-score are usually the best, indicating full potential of machine learning in medical image processing,
high reliability across metrics. ultimately improving patient outcomes worldwide.

V. REFERENCES

[1] M. J. McGuinness and D. J. O'Connor, "An Introduction to

Medical Image Processing and Analysis," Journal of Digital
Imaging, vol. 33, no. 1, pp. 45-57, 2020.

[2] J. Yang et al., "pydicom: A Python library for handling

DICOM files," Journal of Open Source Software, vol. 4, no.
34, p. 1247, 2019.
[ Fig 3 Random Forest metrics ]
[3] P. G. Shrimpton, "SimpleITK for Medical Imaging: A
Which Result is Best? Guide," SpringerBriefs in Computer Science, 2021.

[4] F. Isensee et al., "nnU-Net: Self-adapting Framework for

Based on typical outcomes: U-Net-Based Medical Image Segmentation," Nature Methods,
vol. 18, no. 2, pp. 203-211, 2021.
 Random Forest tends to outperform k-NN and SVM
in terms of accuracy and overall reliability due to its [5] K. Clark et al., "The Cancer Imaging Archive (TCIA):
ensemble method that reduces variance and bias. Maintaining and Operating a Public Information Repository,"
 SVM may provide similar accuracy to RF on linearly Journal of Digital Imaging, vol. 26, no. 6, pp. 1045-1057,
separable datasets but lacks flexibility for non-linear 2013.
data without advanced kernel techniques.
 k-NN, while effective as a baseline, often lags behind
RF and SVM in both accuracy and precision,
especially on complex datasets.

To identify the best-performing model explicitly for this

dataset, you would compare their actual accuracy scores from
the output, along with precision, recall, and F1-scores.
However, based on the general characteristics of these
classifiers and the structured nature of the breast cancer
dataset, Random Forest is likely the best choice for this
task.

IV. CONCLUSIONS

The application of machine learning classifiers in medical

image processing represents a pivotal advancement in modern
healthcare, offering unparalleled accuracy and efficiency in
diagnosing complex medical conditions. By leveraging
algorithms such as k-Nearest Neighbors (k-NN), Support
Vector Machines (SVM), and Random Forests (RF), we can
process and interpret high-dimensional imaging data with
remarkable precision. Each classifier brings unique strengths:
k-NN is effective for straightforward classification tasks,
SVM excels in handling linearly separable data, and Random
Forest demonstrates robustness and scalability for complex
datasets. These tools reduce the burden on medical
professionals, enhance diagnostic reliability, and enable faster
decision-making, particularly in scenarios involving vast
datasets or subtle abnormalities that are difficult to detect
manually.

While these advancements are transformative, challenges

remain, including the need for high-quality, annotated datasets
and concerns surrounding model interpretability and
integration into clinical workflows. Nevertheless, the synergy
between machine learning and medical imaging is poised to
revolutionize healthcare delivery by fostering personalized
medicine and early disease detection. Future research should
focus on developing explainable AI models, enhancing data