Unsupervised discovery of ischemic stroke phenotypes from multimodal MRI radiomics

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Ischemic stroke lesions exhibit significant heterogeneity.
• Lack of annotated outcomes and limited resources hinder lesion classification.
• Multimodal MRI data presents challenges for unsupervised analysis.

Purpose of the Study:

• To develop an unsupervised, label-independent radiomic pipeline for grouping ischemic stroke lesions.
• To address lesion heterogeneity and resource limitations in stroke imaging.
• To enable biologically meaningful lesion stratification without manual segmentation.

Main Methods:

• Analysis of Fluid Attenuated Inversion Recovery (FLAIR), Apparent Diffusion Coefficient (ADC), and Susceptibility Weighted Imaging (SWI) sequences.
• Lesion identification using percentile-based thresholds and variance-filtered feature selection.
• Dimensionality reduction with Uniform Manifold Approximation and Projection (UMAP) followed by clustering (K-means++, Agglomerative, Spectral).

Main Results:

• UMAP combined with Agglomerative clustering yielded high silhouette scores (0.784 for 3 clusters, 0.778 for 5 clusters).
• Optimal consensus stability was achieved (PAC score of 0.000).
• Kruskal-Wallis testing identified 25-36 significant features, primarily from ADC and SWI sequences, with a 5-cluster model showing finer phenotypic separation.

Conclusions:

• The unsupervised framework enables biologically meaningful lesion stratification.
• The pipeline is scalable and suitable for low-resource environments.
• This approach provides a foundation for future diagnostic and prognostic modeling in stroke imaging.