Parsing disease heterogeneity using normative modelling and Generative Adversarial Networks (GANs)

1867 

Abstract Submission 

Sai Spandana Chintapalli¹, Sindhuja Govindarajan¹, Haochang Shou¹, Hao Huang², Christos Davatzikos¹

¹Centre for Biomedical Image Computing and Analytics, University of Pennsylvania, Philadelphia, PA, ²Department of Radiology, Children’s Hospital of Philadelphia, Philadelphia, PA

Sai Spandana Chintapalli  
Centre for Biomedical Image Computing and Analytics, University of Pennsylvania
Philadelphia, PA

Sindhuja Govindarajan  
Centre for Biomedical Image Computing and Analytics, University of Pennsylvania
Philadelphia, PA

Haochang Shou  
Centre for Biomedical Image Computing and Analytics, University of Pennsylvania
Philadelphia, PA

Hao Huang  
Department of Radiology, Children’s Hospital of Philadelphia
Philadelphia, PA

Christos Davatzikos  
Centre for Biomedical Image Computing and Analytics, University of Pennsylvania
Philadelphia, PA

Neuroimaging studies have documented brain structural and functional heterogeneity in patients with Alzheimer's disease (AD) and other neurological disorders.^1,2,9 This heterogeneity leads to diagnostic and prognostic uncertainty, confounding clinical treatment planning. One way to parse disease heterogeneity is normative modelling, where individual-level deviations in brain measures from a reference sample are computed to infer personalized effects of disease.⁸ Traditional univariate normative modelling techniques like Gaussian process regression (GPR) ignore multivariate interactions between brain measures. On the other hand, multivariate deep-learning based techniques such as adversarial autoencoders (AAE) might have low specificity to disease effects as they are trained solely on the reference sample.⁷ In both cases, the computed deviations might incorporate disease irrelevant effects due to inter-individual brain differences. To overcome this, we propose a Generative Adversarial Network (GAN)⁵ based normative modelling technique that learns to remove disease-related variations from a subject's brain measures while preserving disease unrelated variations. As illustrated in Fig.1, the proposed model synthesizes patient-specific controls, and the deviation of the patient from the synthesized control acts as an image-based biomarker that is sensitive to disease effects and severity.

We adapt the pix2pix GAN⁵ to translate a subject with disease to a corresponding subject without disease. Training such a network requires paired data i.e. neuroimaging derived brain measures of an individual with and without disease. In reality, only one of these conditions can be satisfied i.e. subjects either have disease or do not. Hence, we synthetically simulate patients from a known reference sample of controls and use the pseudo-synthetic patient and real control pairs for GAN training. To implement this method to study neuroanatomical heterogeneity, we select a reference population of 6000 healthy controls from the ISTAGING consortium⁴ without preexisting neurologic disorders. Our neuroanatomical brain measures are the 8 regions of interest (ROI) volumes (the left and right frontal, parietal, occipital and temporal lobe volumes) computed using a multi-atlas segmentation technique.³ To simulate patients, for each control in the reference sample we introduce 10-30% atrophy or expansion in a random combination of ROIs while preserving clinical covariate effects. The model is optimized to translate pseudo-synthetic patients to real controls. During inference, the model synthesizes a control for each real patient, and the difference between the two relates to real disease effects. For performance assessment, we select 200 controls (CN) and 200 AD participants (PT) from an independent⁶ dataset and compute their deviations across the 8 ROI volumes using GAN, GPR, and AAE models (pretrained on the ISTAGING dataset). We then use logistic regression to assess the overall discriminative power of the GAN-derived deviations in AD classification.

We observe that the deviations derived using the GAN model are on average larger than the deviations from the GPR and AAE models for PT. While for CN, the GAN's deviations are on average smaller (Fig.2.a). Larger deviations in PT compared to CN reflects that the deviations capture disease related abnormality in brain measures. Additionally, we note that, GAN's deviations in the 8 ROIs (mean AUC = 0.76) provide a considerable gain over raw ROI volumes (mean AUC = 0.65) in classifying AD participants (Fig.2.b).

GAN-based normative modelling technique introduced here is a useful tool to parse heterogeneity in brain measures at an individual level. We see that self-supervised training of the model using pseudo-synthetically simulated patient data that is agnostic to disease patterns can help detect real disease related effects.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling

Methods Development ¹

Multivariate Approaches

Degenerative Disease

Machine Learning

Other - Generative Adversarial Networks (GANs); Normative Modelling; Disease Heterogeneity; Self-Supervised Learning; Synthetic Data