Multi-scale semi-supervised clustering of brain images: Deriving disease subtypes

Area of Science:

• Neuroimaging analysis
• Computational neuroscience
• Machine learning for healthcare

Background:

• Disease heterogeneity complicates understanding and treating brain disorders.
• Current clustering methods struggle with non-disease-related variations and fixed-scale features.
• Semi-supervised techniques offer improved disease-specific pattern identification.

Purpose of the Study:

• To introduce a novel multi-scale clustering method (MAGIC) for brain disease heterogeneity.
• To derive multi-scale, clinically interpretable features for disease subtype identification.
• To evaluate MAGIC's performance and provide guidance on its application.

Main Methods:

• Developed Multi-scAle heteroGeneity analysIs and Clustering (MAGIC), building on the HYDRA method.
• Employed a double-cyclic optimization procedure for inter-scale-consistent subtype identification.
• Conducted extensive semi-simulated experiments on UK Biobank data and applied MAGIC to Alzheimer's and schizophrenia datasets.

Main Results:

• MAGIC effectively depicts multi-scale disease heterogeneity.
• The method was evaluated on healthy controls and applied to Alzheimer's Disease Neuroimaging Initiative (ADNI) and PHENOM datasets.
• Demonstrated potential and identified challenges in dissecting neuroanatomical heterogeneity.

Conclusions:

• MAGIC offers a novel approach to identifying disease-specific subtypes using multi-scale neuroimaging data.
• The study provides insights into the conditions for successful application of such clustering models.
• Guidance is offered on the appropriate use and interpretation of multi-scale heterogeneity analyses in brain disease research.