Unsupervised brain tumor MRI image grouping using deep autoencoder embeddings and K-means clustering

Abstract The severity assessment of brain tumors is an important aspect of clinical decision-making. However, it often depends on expert interpretation and annotated data. The proposed framework organizes brain tumor MRI images into three latent imaging-pattern groups using deep autoencoder embeddings and K-Means clustering. The organized groups represent unsupervised similarities in imaging characteristics and are not intended as clinically validated tumor grades. The proposed method is compared to the popular hybrid models, such as Principal Component Analysis (PCA) + K-Means and VGG16 + K-Means. The performance of the proposed method and other models is measured by Silhouette score, the Calinski-Harabasz index, the maximum cluster separation and Davies-Bouldin index which measures the smallest intra-cluster similarity. The learned latent feature representations are qualitatively visualized using t-distributed Stochastic Neighbour Embedding (t-SNE), which provides insight into the organization of MRI images within the learned latent feature space. The autoencoder-based method in all cases achieved the best results, particularly, in the highest Silhouette Score (0.1601), the Calinski-Harabasz (CH) Index (444.42), and Davies-Bouldin (DB) Index (1.939). The results indicate that the proposed unsupervised deep learning framework is capable of organizing brain tumor MRI images into meaningful latent-space groups. This may support exploratory analysis in other medical imaging scenarios where labelled data are limited. The future work includes clinical validation using expert annotations or graded datasets. This work primarily supports SDG 3: Good Health and Well-Being by exploring scalable computational approaches for unsupervised brain tumor MRI pattern analysis, which may assist future research in medical image interpretation and decision-support systems. It also contributes to SDG 9: Industry, Innovation and Infrastructure by advancing innovative, data-efficient artificial intelligence techniques for medical imaging applications.