Imaging Consortia: Increasing sample size and Understanding heterogeneity in health and disease

In order to fully leverage the potential of machine learning, but also to obtain a deeper understanding of the heterogeneity of imaging signatures revealed by AI methods, our lab has taken the lead to bring together international consortia under the goal of sharing, harmonizing, and integrating imaging and clinical data from various sites around the USA and worldwide. Examples of 3 such consortia are described below

iSTAGING:  imaging-based SysTem for AGing and NeurodeGenerative diseases

sMRI, rsfMRI (fo a subset), amyloid (for a subset) scans from approximately 32,000 individuals aged 45 and older, participants to 13 studies (many of them multi-site), have been harmonized [1] and integrated into this cohort, aiming to create a brain chart in aging. An example of increasing “expression” of SPARE-AD, an AD imaging signature of Alzheimer’s Disease, is shown below:

The goal of this study is to reduce all these complex imaging signals down into a few dozen dimensions. Each of these dimensions reflects the “expression” of an imaging signature informed by clinically and biologically meaningful measures.

PHENOM:  Psychosis Heterogeneity (evaluated) via dimEnsional NeurOiMaging

We are investigating heterogeneity of imaging signatures in schizophrenia, using machine learning methods and including semi-supervised learning methods [2, 3]. The PHENOM consortium brings people and data together from the USA, the UK, Germany, Netherlands, Spain, Australia, and Brazil, for a total of over 3,000 sMRI scans from patients with chronic and first episode schizophrenia, as well as healthy controls. The first studies of this consortium identified two neuroanatomically very different subtypes of SCZ: approximately 2/3 of the patients had the commonly reported diffuse cortical atrophy, however ~1/3 of the patients had intact brains, with the exception of increased volumes in the striatum, which were not explained by medication data. The former, but not the latter, displayed increased gray matter loss with disease duration [4].

ReSPOND: Radiomics Signatures for PrecisiON Diagnostics in Glioblastoma

Artificial intelligence (AI) and machine learning (ML) methods have begun to reveal that complex imaging patterns can provide individualized biomarkers for diagnosis and prognosis. However, AI methods have been challenged by insufficient training, heterogeneity of imaging protocols across hospitals, and lack of generalization to new patient data. These challenges prompted the development of the ReSPOND  consortium on glioblastoma (GBM) [5]. This collaboration of over 10 institutions, across three continents, is positioned to pool, harmonize and analyze brain MRIs from more than 3,300 de novo GBM patients who underwent the Stupp protocol, in addition to datasets from The Cancer Imaging Archive, TCIA[6]. ReSPOND aims to further develop and test AI-based biomarkers for individualized prediction and prognostication, by moving from single-institution studies to generalized, well-validated predictive biomarkers in the following four areas:

1) Overall and progression-free survival. Prior work has shown that informative pre-operative predictors of patient survival can be constructed using imaging-based ML methods [7](Fig-a). These individualized prognostic biomarkers may assist in targeted enrollment and enrichment of clinical trials. They also have the potential to support patient management by identifying poor-prognosis patients who may benefit from early initiation of alternative (to standard) or additional treatments.

2) Peri-tumoral infiltration. Prior work using ML-based imaging signatures has shown promise for identifying tumor infiltration beyond the visible margins and into peri-tumoral edematous (“bright-FLAIR”) tissue [8] (Fig-b). These imaging signatures have been found to identify tissue that is 10 times more likely to present early recurrence, and hence could help establish aggressive, yet targeted, treatments of GBM via peri-tumoral dose escalation and extensive resection.

3) Pseudoprogression(PsP). ML-based imaging signatures have been found to differentiate between treatment-related pseudoprogression and progressive disease [9, 10].

4) Imaging subtypes of GBM. Unsupervised clustering methods applied to rich imaging features have previously identified three imaging subtypes of GBM with divergent prognosis and molecular compositions. These imaging subtypes could help refine WHO classifications, as they appear to offer prognostic information complementary to IDH1 mutation status [11](Fig-c)

1.            Pomponio, R., et al., Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan. Neuroimage, 2020. 208: p. 116450.

2.            Dong, A., et al., CHIMERA: Clustering of heterogeneous disease effects via distribution matching of imaging patterns. IEEE Trans Med Imaging, 2016. 35(2): p. 612-621.

3.            Varol, E., et al., HYDRA: Revealing heterogeneity of imaging and genetic patterns through a multiple max-margin discriminative analysis framework. Neuroimage, 2017. 145(Pt B): p. 346-364.

4.            Chand, G., et al., Two Distinct Neuroanatomical Subtypes of Schizophrenia Revealed Using Machine Learning. Oxford Press, 2020. In Press.

5.            Davatzikos, C., et al., AI based Prognostic Imaging Biomarkers for Precision Neurooncology: the ReSPOND consortiumNeurooncology: the ReSPOND consortium. In Press, 2020.

6.            Clark, K., et al., The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. Journal of Digital Imaging, 2013. 26(6): p. 1045-1057.

7.            Macyszyn, L., et al., Imaging patterns predict patient survival and molecular subtype in glioblastoma via machine learning techniques. Neuro Oncol, 2016. 18(3): p. 417-25.

8.            Akbari, H., et al., Imaging Surrogates of Infiltration Obtained Via Multiparametric Imaging Pattern Analysis Predict Subsequent Location of Recurrence of Glioblastoma. Neurosurgery, 2016. 78(4): p. 572-80.

9.            Nabil Elshafeey, A.K., Dunia Giniebra Camejo, Srishti Abrol, Islam Hassan, Kamel El Salek, Ahmed Hassan, Ahmed Shaaban, Samuel Bergamaschi, Fanny E Moron, Meng Law, Pascal Zinn, Rivka R Colen, Multicenter study to demonstrate radiomic texture features derived from MR perfusion images of pseudoprogression compared to true progression in glioblastoma patients. Journal of Clinical Oncology 2017. 35(15): p. 2016-2016.

10.          Prasanna, P., et al., DISTINGUISHING RADIATION NECROSIS FROM BRAIN TUMOR RECURRENCE ON ROUTINE MRI: A PRELIMINARY HUMAN-MACHINE READER COMPARISON STUDY Neuro-Oncology, 2016. 18(suppl_6): p. vi139–vi140.

11.          Rathore, S., et al., Radiomic MRI signature reveals three distinct subtypes of glioblastoma with different clinical and molecular characteristics, offering prognostic value beyond IDH1. Nature Scientific Reports, 2018. 8(1): p. 5087.