Unsupervised discovery of ischemic stroke phenotypes from multimodal MRI radiomics

<i>Objective.</i>This study presents a fully unsupervised and label-independent radiomic pipeline designed to group different types of ischemic stroke lesions using multimodal Magnetic Resonance Imaging (MRI) . The aim is to address lesion heterogeneity and the absence of annotated outcomes, particularly in settings with limited resources.<i>Approach</i>. Three MRI sequences were analyzed: Fluid Attenuated Inversion Recovery (FLAIR), Apparent Diffusion Coefficient (ADC), and Susceptibility Weighted Imaging (SWI). Lesion identification was performed using percentile-based thresholds, and feature selection was guided by variance filtering with a minimum threshold of 0.001. The complexity of the data was reduced using Uniform Manifold Approximation and Projection (UMAP). Grouping of the lesions was conducted using K-means++, agglomerative hierarchical clustering with Ward linkage, and spectral clustering with a nearest neighbour affinity matrix. The quality and stability of the identified clusters were rigorously evaluated using established internal validation metrics. Feature significance was determined using Kruskal-Wallis testing with Bonferroni correction.<i>Main results</i>. The combination of UMAP with Agglomerative clustering produced the highest silhouette scores of 0.784 for three clusters and 0.778 for five clusters. Consensus stability was optimal, with a Proportion of Ambiguous Clustering score (PAC) of 0.000. Kruskal-Wallis analysis identified 25 significant features for the three-cluster solution and 36 for the five-cluster solution. The most discriminative features originated from ADC and SWI sequences. The five-cluster model revealed finer phenotypic separation and identified five borderline cases with low silhouette coefficients, indicating transitional lesion patterns.<i>Significance</i>. This unsupervised framework enables biologically meaningful lesion stratification without reliance on manual segmentation or outcome labels. It offers a scalable solution for deployment in low-resource environments and provides a robust foundation for future diagnostic and prognostic modelling in stroke imaging.