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

Abstract

Disease heterogeneity is a significant obstacle to understanding pathological processes and delivering precision diagnostics and treatment. Clustering methods have gained popularity in 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 at a basis of feature sets pre- defined at a fixed scale or scales (e.g, an 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 drive inter-scale-consistent disease subtypes or neuroanatomical dimensions effectively. More importantly, to fill in the gap of understanding under what conditions 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 real imaging data of Alzheimers disease (ADNI, N=1728) to demonstrate its potential and challenges in dissecting the neuroanatomical heterogeneity of brain diseases. Taken together, we aim to provide guidelines on 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.

Surreal-GAN:Semi-Supervised Representation Learning via GAN for uncovering heterogeneous disease-related imaging patterns

“…The family of DNMF methods capture discriminant representation by minimizing and maximizing within-class and among-class variations respectively, thus did not capture heterogeneity within the PT class accurately either. Designed for parsing disease heterogeneity, semi-supervised methods focused on clustering patients into hard categorical subtypes (Varol et al (2016), Dong et al (2015), Wen et al (2021), Yang et al (2021)). With different end points, these methods (including Smile-GAN) can not be directly compared with Surreal-GAN.…”

“…Notably, both methods applied directly in the PT domain, confront main limitations in avoiding potential disease-unrelated brain variations. In contrast, semi-supervised clustering methods (Varol et al (2016), Dong et al (2015), Wen et al (2021), Yang et al (2021)) were proposed to cluster patients via the patterns or transformations between the reference group (CN) and the target patient group (PT). A recent proposed deep learning-based semi-supervised clustering method, termed Smile-GAN (Yang et al (2021)), achieved better clustering performance by learning multiple mappings from CN to PT with an inverse clustering function for both regularization and clustering.…”

Genetic, Clinical Underpinnings of Brain Change Along Two Neuroanatomical Dimensions of Clinically-defined Alzheimer’s Disease

“…Surreal-GAN 3 dissects underlying disease-related heterogeneity via a deep representation learning approach under the principle of semi-supervised clustering. Semi-supervised clustering 1,31 seeks the “ 1-to-k ” mapping between the reference healthy control group and the patient group, thereby teasing out clusters or subtypes driven by different pathological trajectories instead of global similarity/dissimilarity in data. Refer to Supplementary eMethod 1 ’s schematic figure for the intuition of deep semi-supervised learning.…”

“…In particular, AI has been applied to MRI to disentangle the neuroanatomical heterogeneity of AD with categorical disease subtypes. 1,2,31 The genetic underpinnings 32,33 of this neuroanatomical heterogeneity in AD are also complex and heterogeneous. The most recent large-scale genome-wide association study 32 (GWAS: 111,326 AD vs. 677,633 controls) has identified 75 genomic loci, including APOE genes, associated with AD.…”