Multi-scale semi-supervised clustering of brain images: deriving disease subtypes
Junhao Wen, Erdem Varol, Aristeidis Sotiras, Zhijian Yang, Ganesh B. Chand, Güray Erus, Haochang Shou, Ahmed Abdulkadir, Gyujoon Hwang, Dominic Dwyer, Alessandro Pigoni, Paola Dazzan, René S. Kahn, Hugo G. Schnack, Marcus V. Zanetti, Eva Meisenzahl, Geraldo F. Busatto, Benedicto Crespo‐Facorro, Rafael Romero-García, Christos Pantelis, Stephen J. Wood, Chuanjun Zhuo, Russell T. Shinohara, Yong Fan, Ruben C. Gur, Raquel E. Gur, Theodore D. Satterthwaite, Nikolaos Koutsouleris, Daniel H. Wolf, Christos Davatzikos
Disease heterogeneity is a significant obstacle to understanding pathological processes and delivering precision diagnostics and treatment. Clustering methods have gained popularity for stratifying patients into subpopulations (i.e., subtypes) of brain diseases using imaging data. However, unsupervised clustering approaches are often confounded by anatomical and functional variations not related to a disease or pathology of interest. Semi-supervised clustering techniques have been proposed to overcome this and, therefore, capture disease-specific patterns more effectively. An additional limitation of both unsupervised and semi-supervised conventional machine learning methods is that they typically model, learn and infer from data using a basis of feature sets pre-defined at a fixed anatomical or functional scale. Herein we propose a novel method, “Multi-scAle heteroGeneity analysIs and Clustering” (MAGIC), to depict the multi-scale presentation of disease heterogeneity, which builds on a previously proposed semi-supervised clustering method, HYDRA. It derives multi-scale and clinically interpretable feature representations and exploits a double-cyclic optimization procedure to effectively drive identification of inter-scale-consistent disease subtypes.
Introduction
Brain diseases like Alzheimer’s and Schizophrenia are notoriously heterogeneous; two patients with the same diagnosis can have vastly different patterns of brain atrophy and clinical symptoms. This makes "one-size-fits-all" treatments ineffective. While clustering algorithms aim to group patients into subtypes, they face two major hurdles: confounding variation (e.g., age, sex) and the scale problem (global vs. focal effects).
To address these, we introduce MAGIC (Multi-scAle heteroGeneity analysIs and Clustering), which builds on the HYDRA framework to identify distinct disease subtypes across multiple anatomical scales simultaneously.
Methods
The MAGIC architecture consists of three core pillars:
Multi-scale Feature Extraction
MAGIC uses orthogonal projective Non-negative Matrix Factorization (opNMF) to decompose brain images into components ranging from coarse (large brain areas) to fine (small, specific structures). This ensures that both global and focal pathological signatures are captured.
Figure 1: Multi-scale feature extraction using opNMF, decomposing voxel data (X) into components (C) and image-level loadings (L).
Unlike unsupervised methods, MAGIC uses a healthy control (CN) group as a reference. It identifies subtypes by determining the specific ways a patient's brain differs from the normative healthy population, effectively filtering out non-pathological variance.
Double-Cyclic Optimization
The "secret sauce" of MAGIC is its double-cyclic optimization procedure. This algorithm iteratively transfers and fine-tunes subtype assignments across different scales, ensuring that the identified subtypes are consistent whether looking at the brain through a wide-angle lens or a microscope.
Figure 2: The double-cyclic optimization process (Cycle 1: polytope transfer; Cycle 2: polytope initialization) ensures inter-scale consistency.
The study utilized three major neuroimaging cohorts:
UK Biobank (UKBB): Data from 4,403 healthy controls was used for "semi-simulated" experiments, where artificial atrophy was imposed to test MAGIC's recovery of ground-truth subtypes.
ADNI: 1,728 participants with Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI).
PHENOM: 1,166 participants used to study Schizophrenia (SCZ).
Key Results
Benchmark Performance: In simulated tests, MAGIC significantly outperformed standard tools like K-means and GMM, especially when disease effects were subtle.
Schizophrenia (SCZ): MAGIC identified two robust subtypes: one with widespread gray matter loss and another with increased basal ganglia volume.
Alzheimer’s (AD): The model found four subtypes, including a "typical" AD pattern and a "normal-like" pattern with minimal structural changes despite clinical symptoms.
Classification Boost: MAGIC-derived features achieved a patient-vs-control classification accuracy of 85%, outperforming linear SVMs (82%).
Conclusion
MAGIC provides a robust framework for uncovering the multi-scale presentation of disease heterogeneity. While challenges remain regarding longitudinal progression and sample size sensitivity, the method offers a significant step toward precision diagnostics. The code is publicly available at GitHub.
Summary updated
Abstract
Abstract Disease heterogeneity is a significant obstacle to understanding pathological processes and delivering precision diagnostics and treatment. Clustering methods have gained popularity for stratifying patients into subpopulations (i.e., subtypes) of brain diseases using imaging data. However, unsupervised clustering approaches are often confounded by anatomical and functional variations not related to a disease or pathology of interest. Semi-supervised clustering techniques have been proposed to overcome this and, therefore, capture disease-specific patterns more effectively. An additional limitation of both unsupervised and semi-supervised conventional machine learning methods is that they typically model, learn and infer from data using a basis of feature sets pre-defined at a fixed anatomical or functional scale (e.g., atlas-based regions of interest). Herein we propose a novel method, “Multi-scAle heteroGeneity analysIs and Clustering” (MAGIC), to depict the multi-scale presentation of disease heterogeneity, which builds on a previously proposed semi-supervised clustering method, HYDRA. It derives multi-scale and clinically interpretable feature representations and exploits a double-cyclic optimization procedure to effectively drive identification of inter-scale-consistent disease subtypes. More importantly, to understand the conditions under which the clustering model can estimate true heterogeneity related to diseases, we conducted extensive and systematic semi-simulated experiments to evaluate the proposed method on a sizeable healthy control sample from the UK Biobank (N=4403). We then applied MAGIC to imaging data from Alzheimer’s disease (ADNI, N =1728) and schizophrenia (PHENOM, N =1166) patients to demonstrate its potential and challenges in dissecting the neuroanatomical heterogeneity of common brain diseases. Taken together, we aim to provide guidance regarding when such analyses can succeed or should be taken with caution. The code of the proposed method is publicly available at https://github.com/anbai106/MAGIC . Highlights We propose a novel multi-scale semi-supervised clustering method, termed MAGIC, to disentangle the heterogeneity of brain diseases. We perform extensive semi-simulated experiments on large control samples (UK Biobank, N =4403) to precisely quantify performance under various conditions, including varying degrees of brain atrophy, different levels of heterogeneity, overlapping disease subtypes, class imbalance, and varying sample sizes. We apply MAGIC to MCI and Alzheimer’s disease (ADNI, N =1728) and schizophrenia (PHENOM, N =1166) patients to dissect their neuroanatomical heterogeneity, providing guidance regarding the use of the semi-simulated experiments to validate the subtypes found in actual clinical applications. Graphical abstract
BibTeX
@article{wen2022multi,
author = {Junhao Wen and Erdem Varol and Aristeidis Sotiras and Zhijian Yang and Ganesh B. Chand and Güray Erus and Haochang Shou and Ahmed Abdulkadir and Gyujoon Hwang and Dominic Dwyer and Alessandro Pigoni and Paola Dazzan and René S. Kahn and Hugo G. Schnack and Marcus V. Zanetti and Eva Meisenzahl and Geraldo F. Busatto and Benedicto Crespo‐Facorro and Rafael Romero-García and Christos Pantelis and Stephen J. Wood and Chuanjun Zhuo and Russell T. Shinohara and Yong Fan and Ruben C. Gur and Raquel E. Gur and Theodore D. Satterthwaite and Nikolaos Koutsouleris and Daniel H. Wolf and Christos Davatzikos and Christos Davatzikos},
title = {{Multi-scale semi-supervised clustering of brain images: deriving disease subtypes}},
journal = {Medical Image Analysis},
year = {2022},
doi = {10.1016/j.media.2021.102304},
}