Brain tumours are seen as critical health challenges that require accurate and early detection for effective treatment. An evaluation of two unsupervised machine learning models Gaussian mixture Fuzzy C Means ensemble model and Spectral Clustering to see the potential use case for potential medical applications. Both models evaluated 800 samples dataset of MRI images using 80 percent of data as training with 20 percent left as testing. The Gaussian Mixture - Fuzzy C Means showed a perfect prediction of all gliomas as opposed to Spectral encountering a catastrophic failure in the “No Tumour” class where all were misclassified as other types of tumours.
Table of Contents
- Abstract
- Introduction
- Literature Review
- Methodology
- MRI Brain Dataset and Preprocessing
- Feature Extraction
- Intensity features
- Gradient Texture
- Morphological Features
- Frequency Features
- Statistics
- Edge Detection
- Histogram
- Gaussian Mixture Model and Fuzzy C Means Ensemble Model (GMM-FCM)
- Spectral Clustering Model including Multiscale Analysis
- Alignment Of Clusters and Evaluation
- Results and Discussion
- Model Performance
- Models Confusion Matrix
- ROC-AUC Comparison
- Conclusion
- References
Objective & Themes
This research aims to compare the potential utility of two unsupervised machine learning models, Gaussian Mixture Fuzzy C Means (GMM-FCM) and Spectral Clustering, for the detection and classification of brain tumours in MRI images. The study evaluates their performance on a dataset of brain MRI images to identify the most suitable model for clinical applications.
- Unsupervised Machine Learning for Medical Imaging
- Brain Tumour Detection and Classification
- Performance Comparison of GMM-FCM and Spectral Clustering
- MRI Image Analysis and Feature Extraction
- Model Evaluation Metrics (Accuracy, ROC-AUC, Confusion Matrix)
- Clinical Applicability and Limitations of ML Models
Excerpt from the Book
Gaussian Mixture Model and Fuzzy C Means Ensemble Model (GMM-FCM)
The GMM-FCM ensemble model is a hybrid approach of probabilistic and geometric clustering. The Gaussian Mixture component will assume the datapoints that are created from distributions of multiple Gaussian models with each one representing a tumor class.
The Fuzzy C Means Component introduces soft clustering as each image is given a degree of membership to all clusters as opposed to belonging to only one cluster. This is valuable in imaging where brain tumours can be unclear with mixed characteristics.
The model functions by updating assigned clusters through two processes. The expectation will compute the membership probabilities using GMM's likelihood and FCM's distance assignments. Maximization will update the centres of each cluster and mixing weights based on these connections. Hybrid parameter (Gamma =6) controls how much will be the model balance between GMM and FCM with this parameter favouring GMM and using FCM to outliers.
For initialization, K-means++ seeding was used to establish the starting positions for the cluster centres to reduce sensitivity. An adaptive fussiness parameter was added to adjust the cluster boundaries automatically. To predict new test images, k-nearest neighbour weighted votes based on inverse distance.
Summary of Chapters
Abstract: This section introduces brain tumours as critical health challenges, detailing the evaluation of two unsupervised machine learning models (GMM-FCM and Spectral Clustering) for MRI brain tumour detection and concluding that GMM-FCM showed superior potential for clinical applications.
Introduction: The introduction highlights the severe impact of brain tumours and the growing role of machine learning in aiding early diagnosis, outlining the study's objective to compare GMM-FCM and Spectral models in classifying brain tumours from a given MRI dataset.
Literature Review: This chapter provides background on tumour formation, the need for high-accuracy automated diagnosis, and the utility of quantitative medical image analysis, discussing previous studies that utilized Gaussian Mixture Models, Spectral Clustering, and Fuzzy C Means in similar contexts.
Methodology: The methodology details the experimental setup, including the use of an MRI brain dataset from Kaggle, the preprocessing steps, the comprehensive feature extraction process, and the specific implementation of the GMM-FCM and Spectral Clustering models, along with their evaluation strategy.
Results and Discussion: This chapter presents the performance metrics of both models, including accuracy, precision, recall, F1-score, and ROC-AUC, highlighting GMM-FCM's superior classification capabilities, especially in glioma detection, and Spectral Clustering's significant failures in classifying "No Tumour" cases.
Conclusion: The conclusion reiterates GMM-FCM's greater potential over Spectral Clustering for brain tumour classification due to its superior performance across all metrics and perfect prediction in the "Glioma" class, while also discussing the limitations of the study and suggesting directions for future research.
References: This section lists all academic sources, research papers, and online resources cited throughout the document, providing the bibliography for the study.
Keywords
Artificial Intelligence, Machine Learning, Brain Tumour Detection, MRI, Gaussian Mixture Fuzzy C Means (GMM-FCM), Spectral Clustering, Unsupervised Learning, Image Segmentation, Feature Extraction, Classification, Performance Metrics, ROC-AUC, Confusion Matrix, Glioma, Meningioma, Pituitary.
Frequently Asked Questions
What is this work fundamentally about?
This work fundamentally compares two unsupervised machine learning models, Gaussian Mixture Fuzzy C Means (GMM-FCM) and Spectral Clustering, for their effectiveness in detecting and classifying brain tumours from MRI images.
What are the central thematic areas?
The central thematic areas include artificial intelligence, machine learning, medical image analysis (specifically MRI brain scans), unsupervised clustering algorithms, and the application of these technologies for disease diagnosis and detection in healthcare.
What is the primary goal or research question?
The primary goal is to evaluate the potential use cases of GMM-FCM and Spectral Clustering in MRI brain tumour detection and to determine which model offers better accuracy and reliability for potential clinical applications.
Which scientific method is used?
The scientific method involves employing and comparing two unsupervised clustering algorithms (GMM-FCM and Spectral Clustering) on a dataset of MRI brain images, followed by quantitative evaluation using various performance metrics such as accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices.
What is covered in the main part?
The main part covers the literature review on existing methods, the detailed methodology including data preprocessing, feature extraction, the specifics of the GMM-FCM and Spectral Clustering models, and the comprehensive results and discussion of their performance comparison.
Which keywords characterize the work?
The work is characterized by keywords such as Artificial Intelligence, Machine Learning, Brain Tumour Detection, MRI, GMM-FCM, Spectral Clustering, Unsupervised Learning, Image Segmentation, and Performance Metrics.
Why were unsupervised models chosen for this study?
Unsupervised models were chosen because they can handle unlabelled data and autonomously adapt to patterns, making them useful for detecting anomalies without prior knowledge, which is relevant for analyzing brain MRI images without pre-existing labels.
What was the most significant failure of the Spectral Clustering model?
The most significant failure of the Spectral Clustering model was its catastrophic inability to correctly classify "No Tumour" samples, resulting in 0% correct predictions for healthy brain tissue, where all healthy samples were misclassified as various tumour types.
What are the main limitations identified in this study?
The main limitations include a restricted dataset size of 100 images per class, which limits the diversity of scanner types, and a reliance on pixel-level and basic features without the use of advanced, domain-specific engineering, potentially restricting both models' performance.
What does the GMM-FCM model offer that makes it more suitable for medical imaging compared to single-cluster models?
The GMM-FCM model introduces soft clustering, assigning a degree of membership to all clusters for each image rather than a single cluster. This is particularly valuable in medical imaging, where brain tumours can have unclear and mixed characteristics, making it difficult to assign them to a single, distinct category.
- Citar trabajo
- Jheirom Pablo (Autor), 2026, A comparison of two machine learning models, Múnich, GRIN Verlag, https://www.grin.com/document/1692018
