A Clustering-based Hybrid Approach for Analyzing High-grade Gliomas Using Radiomic Features

Abstract

Unlabeled data is a significant problem in healthcare and other fields that deal with huge datasets. Unsupervised learning has the potential to be an effective solution in this case. The use of unsupervised algorithms in disease diagnosis has not been widely explored. In this work, we have developed a clustering algorithm to analyze the gliomas using Magnetic Resonance Imaging (MRI) data. Glioma is a severe medical illness that necessitates an accurate and timely diagnosis to establish effective treatment options. We used Pyradiomics to extract radiomic characteristics from MRI scans, which were then fed into a number of clustering methods, with cluster fitness assessed using primary assessment metrics. The best clustering algorithm was used as the pre-processor and to train major classification algorithms. In this study, we examined the performance of three prominent clustering algorithms, with agglomerative clustering outperforming the others. We achieved 0.83 Silhouette Coefficient, 0.21 Davies-Bouldin Index, and 323.22 Calinski-Harabasz Index values using aggregative clustering using Pyradiomics features. The decision tree strategy outperformed all classification methods, achieving 99.54% accuracy when clustering was applied to preprocess the data before classification. The proposed work has considerable potential for faster and more accurate analysis of medical image problems, especially in gliomas.