Genetic and Clinical Correlates of AI-Based Brain Aging Patterns in Cognitively Unimpaired Individuals

Importance

Brain aging elicits complex neuroanatomical changes influenced by multiple age-related pathologies. Understanding the heterogeneity of structural brain changes in aging may provide insights into preclinical stages of neurodegenerative diseases.

Objective

To derive subgroups with common patterns of variation in participants without diagnosed cognitive impairment (WODCI) in a data-driven manner and relate them to genetics, biomedical measures, and cognitive decline trajectories.

Design, Setting, and Participants

Data acquisition for this cohort study was performed from 1999 to 2020. Data consolidation and harmonization were conducted from July 2017 to July 2021. Age-specific subgroups of structural brain measures were modeled in 4 decade-long intervals spanning ages 45 to 85 years using a deep learning, semisupervised clustering method leveraging generative adversarial networks. Data were analyzed from July 2021 to February 2023 and were drawn from the Imaging-Based Coordinate System for Aging and Neurodegenerative Diseases (iSTAGING) international consortium. Individuals WODCI at baseline spanning ages 45 to 85 years were included, with greater than 50 000 data time points.

Exposures

Individuals WODCI at baseline scan.

Main Outcomes and Measures

Three subgroups, consistent across decades, were identified within the WODCI population. Associations with genetics, cardiovascular risk factors (CVRFs), amyloid β (Aβ), and future cognitive decline were assessed.

Results

In a sample of 27 402 individuals (mean [SD] age, 63.0 [8.3] years; 15 146 female [55%]) WODCI, 3 subgroups were identified in contrast with the reference group: a typical aging subgroup, A1, with a specific pattern of modest atrophy and white matter hyperintensity (WMH) load, and 2 accelerated aging subgroups, A2 and A3, with characteristics that were more distinct at age 65 years and older. A2 was associated with hypertension, WMH, and vascular disease-related genetic variants and was enriched for Aβ positivity (ages ≥65 years) and apolipoprotein E (APOE) ε4 carriers. A3 showed severe, widespread atrophy, moderate presence of CVRFs, and greater cognitive decline. Genetic variants associated with A1 were protective for WMH (rs7209235: mean [SD] B = -0.07 [0.01]; P value = 2.31 × 10-9) and Alzheimer disease (rs72932727: mean [SD] B = 0.1 [0.02]; P value = 6.49 × 10-9), whereas the converse was observed for A2 (rs7209235: mean [SD] B = 0.1 [0.01]; P value = 1.73 × 10-15 and rs72932727: mean [SD] B = -0.09 [0.02]; P value = 4.05 × 10-7, respectively); variants in A3 were associated with regional atrophy (rs167684: mean [SD] B = 0.08 [0.01]; P value = 7.22 × 10-12) and white matter integrity measures (rs1636250: mean [SD] B = 0.06 [0.01]; P value = 4.90 × 10-7).

Conclusions and Relevance

The 3 subgroups showed distinct associations with CVRFs, genetics, and subsequent cognitive decline. These subgroups likely reflect multiple underlying neuropathologic processes and affect susceptibility to Alzheimer disease, paving pathways toward patient stratification at early asymptomatic stages and promoting precision medicine in clinical trials and health care.

The Genetic Architecture of Biological Age in Nine Human Organ Systems

Understanding the genetic basis of biological aging in multi-organ systems is vital for elucidating age-related disease mechanisms and identifying therapeutic interventions. This study characterized the genetic architecture of the biological age gap (BAG) across nine human organ systems in 377,028 individuals of European ancestry from the UK Biobank. We discovered 393 genomic loci-BAG pairs (P-value<5×10-8) linked to the brain, eye, cardiovascular, hepatic, immune, metabolic, musculoskeletal, pulmonary, and renal systems. We observed BAG-organ specificity and inter-organ connections. Genetic variants associated with the nine BAGs are predominantly specific to the respective organ system while exerting pleiotropic effects on traits linked to multiple organ systems. A gene-drug-disease network confirmed the involvement of the metabolic BAG-associated genes in drugs targeting various metabolic disorders. Genetic correlation analyses supported Cheverud's Conjecture1 - the genetic correlation between BAGs mirrors their phenotypic correlation. A causal network revealed potential causal effects linking chronic diseases (e.g., Alzheimer's disease), body weight, and sleep duration to the BAG of multiple organ systems. Our findings shed light on promising therapeutic interventions to enhance human organ health within a complex multi-organ network, including lifestyle modifications and potential drug repositioning strategies for treating chronic diseases. All results are publicly available at https://labs-laboratory.com/medicine.

The genetic architecture of multimodal human brain age

The complex biological mechanisms underlying human brain aging remain incompletely understood. This study investigated the genetic architecture of three brain age gaps (BAG) derived from gray matter volume (GM-BAG), white matter microstructure (WM-BAG), and functional connectivity (FC-BAG). We identified sixteen genomic loci that reached genome-wide significance (P-value < 5×10-8). A gene-drug-disease network highlighted genes linked to GM-BAG for treating neurodegenerative and neuropsychiatric disorders and WM-BAG genes for cancer therapy. GM-BAG displayed the most pronounced heritability enrichment in genetic variants within conserved regions. Oligodendrocytes and astrocytes, but not neurons, exhibited notable heritability enrichment in WM and FC-BAG, respectively. Mendelian randomization identified potential causal effects of several chronic diseases on brain aging, such as type 2 diabetes on GM-BAG and AD on WM-BAG. Our results provide insights into the genetics of human brain aging, with clinical implications for potential lifestyle and therapeutic interventions. All results are publicly available at https://labs.loni.usc.edu/medicine .

Integrating imaging and genomic data for the discovery of distinct glioblastoma subtypes: a joint learning approach

Glioblastoma is a highly heterogeneous disease, with variations observed at both phenotypical and molecular levels. Personalized therapies would be facilitated by non-invasive in vivo approaches for characterizing this heterogeneity. In this study, we developed unsupervised joint machine learning between radiomic and genomic data, thereby identifying distinct glioblastoma subtypes. A retrospective cohort of 571 IDH-wildtype glioblastoma patients were included in the study, and pre-operative multi-parametric MRI scans and targeted next-generation sequencing (NGS) data were collected. L21-norm minimization was used to select a subset of 12 radiomic features from the MRI scans, and 13 key driver genes from the five main signal pathways most affected in glioblastoma were selected from the genomic data. Subtypes were identified using a joint learning approach called Anchor-based Partial Multi-modal Clustering on both radiomic and genomic modalities. Kaplan-Meier analysis identified three distinct glioblastoma subtypes: high-risk, medium-risk, and low-risk, based on overall survival outcome (p < 0.05, log-rank test; Hazard Ratio = 1.64, 95% CI 1.17-2.31, Cox proportional hazard model on high-risk and low-risk subtypes). The three subtypes displayed different phenotypical and molecular characteristics in terms of imaging histogram, co-occurrence of genes, and correlation between the two modalities. Our findings demonstrate the synergistic value of integrated radiomic signatures and molecular characteristics for glioblastoma subtyping. Joint learning on both modalities can aid in better understanding the molecular basis of phenotypical signatures of glioblastoma, and provide insights into the biological underpinnings of tumor formation and progression.

Gene-SGAN: discovering disease subtypes with imaging and genetic signatures via multi-view weakly-supervised deep clustering

Disease heterogeneity has been a critical challenge for precision diagnosis and treatment, especially in neurologic and neuropsychiatric diseases. Many diseases can display multiple distinct brain phenotypes across individuals, potentially reflecting disease subtypes that can be captured using MRI and machine learning methods. However, biological interpretability and treatment relevance are limited if the derived subtypes are not associated with genetic drivers or susceptibility factors. Herein, we describe Gene-SGAN - a multi-view, weakly-supervised deep clustering method - which dissects disease heterogeneity by jointly considering phenotypic and genetic data, thereby conferring genetic correlations to the disease subtypes and associated endophenotypic signatures. We first validate the generalizability, interpretability, and robustness of Gene-SGAN in semi-synthetic experiments. We then demonstrate its application to real multi-site datasets from 28,858 individuals, deriving subtypes of Alzheimer's disease and brain endophenotypes associated with hypertension, from MRI and single nucleotide polymorphism data. Derived brain phenotypes displayed significant differences in neuroanatomical patterns, genetic determinants, biological and clinical biomarkers, indicating potentially distinct underlying neuropathologic processes, genetic drivers, and susceptibility factors. Overall, Gene-SGAN is broadly applicable to disease subtyping and endophenotype discovery, and is herein tested on disease-related, genetically-associated neuroimaging phenotypes.

Five dominant dimensions of brain aging are identified via deep learning: associations with clinical, lifestyle, and genetic measures

Brain aging is a complex process influenced by various lifestyle, environmental, and genetic factors, as well as by age-related and often co-existing pathologies. MRI and, more recently, AI methods have been instrumental in understanding the neuroanatomical changes that occur during aging in large and diverse populations. However, the multiplicity and mutual overlap of both pathologic processes and affected brain regions make it difficult to precisely characterize the underlying neurodegenerative profile of an individual from an MRI scan. Herein, we leverage a state-of-the art deep representation learning method, Surreal-GAN, and present both methodological advances and extensive experimental results that allow us to elucidate the heterogeneity of brain aging in a large and diverse cohort of 49,482 individuals from 11 studies. Five dominant patterns of neurodegeneration were identified and quantified for each individual by their respective (herein referred to as) R-indices. Significant associations between R-indices and distinct biomedical, lifestyle, and genetic factors provide insights into the etiology of observed variances. Furthermore, baseline R-indices showed predictive value for disease progression and mortality. These five R-indices contribute to MRI-based precision diagnostics, prognostication, and may inform stratification into clinical trials.

Genomic loci influence patterns of structural covariance in the human brain

Normal and pathologic neurobiological processes influence brain morphology in coordinated ways that give rise to patterns of structural covariance (PSC) across brain regions and individuals during brain aging and diseases. The genetic underpinnings of these patterns remain largely unknown. We apply a stochastic multivariate factorization method to a diverse population of 50,699 individuals (12 studies and 130 sites) and derive data-driven, multi-scale PSCs of regional brain size. PSCs were significantly correlated with 915 genomic loci in the discovery set, 617 of which are newly identified, and 72% were independently replicated. Key pathways influencing PSCs involve reelin signaling, apoptosis, neurogenesis, and appendage development, while pathways of breast cancer indicate potential interplays between brain metastasis and PSCs associated with neurodegeneration and dementia. Using support vector machines, multi-scale PSCs effectively derive imaging signatures of several brain diseases. Our results elucidate genetic and biological underpinnings that influence structural covariance patterns in the human brain.

The Genetic Architecture of Multimodal Human Brain Age

The complex biological mechanisms underlying human brain aging remain incompletely understood, involving multiple body organs and chronic diseases. In this study, we used multimodal magnetic resonance imaging and artificial intelligence to examine the genetic architecture of the brain age gap (BAG) derived from gray matter volume (GM-BAG, N=31,557 European ancestry), white matter microstructure (WM-BAG, N=31,674), and functional connectivity (FC-BAG, N=32,017). We identified sixteen genomic loci that reached genome-wide significance (P-value<5×10-8). A gene-drug-disease network highlighted genes linked to GM-BAG for treating neurodegenerative and neuropsychiatric disorders and WM-BAG genes for cancer therapy. GM-BAG showed the highest heritability enrichment for genetic variants in conserved regions, whereas WM-BAG exhibited the highest heritability enrichment in the 5' untranslated regions; oligodendrocytes and astrocytes, but not neurons, showed significant heritability enrichment in WM and FC-BAG, respectively. Mendelian randomization identified potential causal effects of several exposure variables on brain aging, such as type 2 diabetes on GM-BAG (odds ratio=1.05 [1.01, 1.09], P-value=1.96×10-2) and AD on WM-BAG (odds ratio=1.04 [1.02, 1.05], P-value=7.18×10-5). Overall, our results provide valuable insights into the genetics of human brain aging, with clinical implications for potential lifestyle and therapeutic interventions. All results are publicly available at the MEDICINE knowledge portal: https://labs.loni.usc.edu/medicine.

Association of brain microbleeds with risk factors, cognition, and MRI markers in MESA

INTRODUCTION

Little is known about the epidemiology of brain microbleeds in racially/ethnically diverse populations.

METHODS

In the Multi-Ethnic Study of Atherosclerosis, brain microbleeds were identified from 3T magnetic resonance imaging susceptibility-weighted imaging sequences using deep learning models followed by radiologist review.

RESULTS

Among 1016 participants without prior stroke (25% Black, 15% Chinese, 19% Hispanic, 41% White, mean age 72), microbleed prevalence was 20% at age 60 to 64.9 and 45% at ≥85 years. Deep microbleeds were associated with older age, hypertension, higher body mass index, and atrial fibrillation, and lobar microbleeds with male sex and atrial fibrillation. Overall, microbleeds were associated with greater white matter hyperintensity volume and lower total white matter fractional anisotropy.

DISCUSSION

Results suggest differing associations for lobar versus deep locations. Sensitive microbleed quantification will facilitate future longitudinal studies of their potential role as an early indicator of vascular pathology.

Multiscale functional connectivity patterns of the aging brain learned from harmonized rsfMRI data of the multi-cohort iSTAGING study

To learn multiscale functional connectivity patterns of the aging brain, we built a brain age prediction model of functional connectivity measures at seven scales on a large fMRI dataset, consisting of resting-state fMRI scans of 4186 individuals with a wide age range (22 to 97 years, with an average of 63) from five cohorts. We computed multiscale functional connectivity measures of individual subjects using a personalized functional network computational method, harmonized the functional connectivity measures of subjects from multiple datasets in order to build a functional brain age model, and finally evaluated how functional brain age gap correlated with cognitive measures of individual subjects. Our study has revealed that functional connectivity measures at multiple scales were more informative than those at any single scale for the brain age prediction, the data harmonization significantly improved the brain age prediction performance, and the data harmonization in the functional connectivity measures' tangent space worked better than in their original space. Moreover, brain age gap scores of individual subjects derived from the brain age prediction model were significantly correlated with clinical and cognitive measures. Overall, these results demonstrated that multiscale functional connectivity patterns learned from a large-scale multi-site rsfMRI dataset were informative for characterizing the aging brain and the derived brain age gap was associated with cognitive and clinical measures.

Gene-SGAN: a method for discovering disease subtypes with imaging and genetic signatures via multi-view weakly-supervised deep clustering

Disease heterogeneity has been a critical challenge for precision diagnosis and treatment, especially in neurologic and neuropsychiatric diseases. Many diseases can display multiple distinct brain phenotypes across individuals, potentially reflecting disease subtypes that can be captured using MRI and machine learning methods. However, biological interpretability and treatment relevance are limited if the derived subtypes are not associated with genetic drivers or susceptibility factors. Herein, we describe Gene-SGAN - a multi-view, weakly-supervised deep clustering method - which dissects disease heterogeneity by jointly considering phenotypic and genetic data, thereby conferring genetic correlations to the disease subtypes and associated endophenotypic signatures. We first validate the generalizability, interpretability, and robustness of Gene-SGAN in semi-synthetic experiments. We then demonstrate its application to real multi-site datasets from 28,858 individuals, deriving subtypes of Alzheimer's disease and brain endophenotypes associated with hypertension, from MRI and SNP data. Derived brain phenotypes displayed significant differences in neuroanatomical patterns, genetic determinants, biological and clinical biomarkers, indicating potentially distinct underlying neuropathologic processes, genetic drivers, and susceptibility factors. Overall, Gene-SGAN is broadly applicable to disease subtyping and endophenotype discovery, and is herein tested on disease-related, genetically-driven neuroimaging phenotypes.

NIMG-82. RADIOGENOMIC SIGNATURES OF KEY DRIVER GENES IN GBM REVEAL MOLECULAR HETEROGENEITY OF THE TUMOR MICROENVIRONMENT LINKED TO SPATIAL DISTRIBUTION: IMPACT ON THE TRAJECTORY OF GLIOMA EVOLUTION

Somatic genomic alterations acquired during GBM growth enhance adaptation of tumor cells to their microenvironment and give rise to molecular heterogeneity. Radiogenomics could facilitate exploration of the underlying pathobiology of tumor growth in specific microenvironments and thereby, promote precision medicine for the patients. We derived radiogenomic signatures of key driver genes and evaluated molecular compositions of tumor groups with predisposition to specific brain regions. Pre-operative multiparametric conventional MRI scans of 357 IDH-wildtype GBM patients with available targeted NGS data were jointly segmented and registered into a common template. We constructed spatial distribution atlases for tumors harboring mutations in driver genes and identified four distinct groups of tumor locations with predilection to the left frontal cingulate region (Group1), right temporal (Group2), right parietal (Group3), and occipital pole (Group4). Evaluation of the differences in molecular features of the tumor groups included: (1) exploring similarities of genomic profiles across all four groups by evaluating cosine similarity metric (CSM) between mutational signatures; (2) quantification of molecular heterogeneity based on Mutant Allele Tumor Heterogeneity (MATH) scores; and (3) inference of the evolutionary trajectories. Groups 1 and 4 were the most different, and Groups 2 and 3 were the most similar tumors, molecularly. The mutational signatures between Groups 1 and 4 revealed a CSM of 0.35. Group1 showed significantly lower MATH score (less heterogeneity) compared to Group4 (p&lt; 0.05). Evaluation of evolutionary patterns suggested NF1 mutation as an early event in Group1, without subsequent gain of function or mutation in EGFR. In contrast, in Group4, EGFR mutations were early events triggering PTEN mutations later in the evolutionary trajectory. Radiogenomic signatures revealed distinct molecular underpinnings for the tumors with predilection towards specific brain regions that may suggest existence of different tumor microenvironments in different brain regions that cause intra- and inter-patient heterogeneity in the molecular tumor composition.

NIMG-22. AN AI-BASED COORDINATE SYSTEM ELUCIDATES RADIOGENOMIC HETEROGENEITY OF GLIOBLASTOMA VIA DEEP LEARNING AND MANIFOLD EMBEDDINGS

PURPOSE

There is evidence that molecular heterogeneity of glioblastoma is associated with heterogeneity of MR imaging signatures. Modern machine learning models, such as deep neural networks, provide a tool for capturing such complex relationships in high-dimensional datasets. This study leverages recent advances in visualizing neural networks to construct a radiogenomics coordinate system whose axes reflect the expression of imaging signatures of genetic mutations commonly found in glioblastoma.

METHODS

Multi-parametric MRI (mpMRI) scans (T1, T1-Gd, T2, T2-FLAIR, DSC, DTI) of 254 subjects with glioblastoma were retrospectively collected. Radiomics features, including histograms, morphologic and textural descriptors, were derived. Genetic markers were obtained through next generation sequencing (NGS) panel. A multi-label classification deep neural network was trained for predicting mutation status in key driver genes, EGFR, PTEN, NF1, TP53 and RB1. We utilized a nonlinear manifold learning method called Intensive Principal Component Analysis (InPCA), to visualize the output probability distributions from the trained model. The first three principal components (PCs) were selected for constructing the coordinate system.

RESULTS

The axes derived from InPCA analysis were associated with molecular pathways known to be implicated in glioblastoma: (1) Increasing values of PC1 were associated with primary involvement of P53 then RB1 then MAPK then RTK/PI3K; (2) Increasing values of PC2 were associated with primary involvement of RTK then RB1/P53/MAPK then PI3K; (3) Increasing values of PC3 were associated with primary involvement of MAPK then RB1/P53/RTK/PI3K. Imaging features significantly associated with each of three PCs (p&lt; 0.05) were identified by Pearson correlation analysis.

CONCLUSION

Deep learning followed by nonlinear manifold embedding identifies a radiogenomics coordinate system spanned by three components which were associated with different molecular pathways of glioblastoma. The heterogeneity of radiogenomic signatures captured by this coordinate system offers in vivo biomarkers of the molecular heterogeneity of glioblastoma.

NIMG-37. JOINT LEARNING OF IMAGING AND GENOMIC DATA REVEALS DISTINCT GLIOBLASTOMA SUBTYPES

PURPOSE

The significant heterogeneity of glioblastoma is typically displayed on both phenotypical and molecular levels. Non-invasive in vivo approaches to characterize this heterogeneity would potentially facilitate personalized therapies. Here we leverage advanced unsupervised machine learning to integrate radiomic imaging features and genomics to identify distinct subtypes of glioblastoma.

METHODS

A retrospective cohort of 571 IDH-wildtype glioblastoma patients were collected with pre-operative multi-parametric MRI (T1, T1CE, T2, T2-FLAIR, DSC, DTI) scans (available in 462/571 patients) and targeted next-generation sequencing (NGS) data (available in 355/571 patients). Radiomic features (n= 971) were extracted from these MRI scans and a subset of 12 features were selected by L21-norm minimization. A total of 14 key driver genes in the 5 main pathways that are most frequently altered in glioblastoma were chosen. Subtypes were identified by a joint learning approach called Anchor-based Partial Multi-modal Clustering (APMC) on both radiomic and genomic modalities.

RESULTS

Three distinct glioblastoma subtypes were discovered by APMC based on 14-dimension NGS data together with 12 selected radiomic features representing characteristics from histograms, shape, and volumetric measures for different tumor sub-regions. The identified subtypes were 1) high-risk; 2) medium-risk; and 3) low-risk, in terms of their overall survival outcome in Kaplan-Meier analysis (p= 5.52e-6, log-rank test; HR= 1.51, 95%CI:1.20-1.74, Cox proportional hazard model). The three subtypes also displayed different molecular characteristics: subtype 1 exhibited increased frequency of mutation in [EGFR, PIK3CA, PTEN, NF1], subtype 3 showed frequently mutated [PDGFRA, ATRX], while subtype 2 did not show significant differences for mutations in any of these genes.

CONCLUSION

Our results revealed the synergistic value of integrated radiomic signatures and molecular characteristics for glioblastoma subtyping. Joint learning on both modalities could help better understand the molecular basis of phenotypical signatures of glioblastoma and further provide insights into the biologic underpinnings of tumor formation and progression.

Schizophrenia Imaging Signatures and Their Associations With Cognition, Psychopathology, and Genetics in the General Population

OBJECTIVE

The prevalence and significance of schizophrenia-related phenotypes at the population level is debated in the literature. Here, the authors assessed whether two recently reported neuroanatomical signatures of schizophrenia-signature 1, with widespread reduction of gray matter volume, and signature 2, with increased striatal volume-could be replicated in an independent schizophrenia sample, and investigated whether expression of these signatures can be detected at the population level and how they relate to cognition, psychosis spectrum symptoms, and schizophrenia genetic risk.

METHODS

This cross-sectional study used an independent schizophrenia-control sample (N=347; ages 16-57 years) for replication of imaging signatures, and then examined two independent population-level data sets: typically developing youths and youths with psychosis spectrum symptoms in the Philadelphia Neurodevelopmental Cohort (N=359; ages 16-23 years) and adults in the UK Biobank study (N=836; ages 44-50 years). The authors quantified signature expression using support-vector machine learning and compared cognition, psychopathology, and polygenic risk between signatures.

RESULTS

Two neuroanatomical signatures of schizophrenia were replicated. Signature 1 but not signature 2 was significantly more common in youths with psychosis spectrum symptoms than in typically developing youths, whereas signature 2 frequency was similar in the two groups. In both youths and adults, signature 1 was associated with worse cognitive performance than signature 2. Compared with adults with neither signature, adults expressing signature 1 had elevated schizophrenia polygenic risk scores, but this was not seen for signature 2.

CONCLUSIONS

The authors successfully replicated two neuroanatomical signatures of schizophrenia and describe their prevalence in population-based samples of youths and adults. They further demonstrated distinct relationships of these signatures with psychosis symptoms, cognition, and genetic risk, potentially reflecting underlying neurobiological vulnerability.

The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics

Glioblastoma is the most common aggressive adult brain tumor. Numerous studies have reported results from either private institutional data or publicly available datasets. However, current public datasets are limited in terms of: a) number of subjects, b) lack of consistent acquisition protocol, c) data quality, or d) accompanying clinical, demographic, and molecular information. Toward alleviating these limitations, we contribute the "University of Pennsylvania Glioblastoma Imaging, Genomics, and Radiomics" (UPenn-GBM) dataset, which describes the currently largest publicly available comprehensive collection of 630 patients diagnosed with de novo glioblastoma. The UPenn-GBM dataset includes (a) advanced multi-parametric magnetic resonance imaging scans acquired during routine clinical practice, at the University of Pennsylvania Health System, (b) accompanying clinical, demographic, and molecular information, (d) perfusion and diffusion derivative volumes, (e) computationally-derived and manually-revised expert annotations of tumor sub-regions, as well as (f) quantitative imaging (also known as radiomic) features corresponding to each of these regions. This collection describes our contribution towards repeatable, reproducible, and comparative quantitative studies leading to new predictive, prognostic, and diagnostic assessments.

Disentangling Alzheimer’s disease neurodegeneration from typical brain ageing using machine learning

Neuroimaging biomarkers that distinguish between changes due to typical brain ageing and Alzheimer’s disease are valuable for determining how much each contributes to cognitive decline. Supervised machine learning models can derive multivariate patterns of brain change related to the two processes, including the Spatial Patterns of Atrophy for Recognition of Alzheimer’s Disease (SPARE-AD) and of Brain Aging (SPARE-BA) scores investigated herein. However, the substantial overlap between brain regions affected in the two processes confounds measuring them independently. We present a methodology, and associated results, towards disentangling the two.

T₁-weighted MRI scans of 4054 participants (48–95 years) with Alzheimer’s disease, mild cognitive impairment (MCI), or cognitively normal (CN) diagnoses from the Imaging-based coordinate SysTem for AGIng and NeurodeGenerative diseases (iSTAGING) consortium were analysed. Multiple sets of SPARE scores were investigated, in order to probe imaging signatures of certain clinically or molecularly defined sub-cohorts. First, a subset of clinical Alzheimer’s disease patients (n = 718) and age- and sex-matched CN adults (n = 718) were selected based purely on clinical diagnoses to train SPARE-BA1 (regression of age using CN individuals) and SPARE-AD1 (classification of CN versus Alzheimer’s disease) models. Second, analogous groups were selected based on clinical and molecular markers to train SPARE-BA2 and SPARE-AD2 models: amyloid-positive Alzheimer’s disease continuum group (n = 718; consisting of amyloid-positive Alzheimer’s disease, amyloid-positive MCI, amyloid- and tau-positive CN individuals) and amyloid-negative CN group (n = 718). Finally, the combined group of the Alzheimer’s disease continuum and amyloid-negative CN individuals was used to train SPARE-BA3 model, with the intention to estimate brain age regardless of Alzheimer’s disease-related brain changes.

The disentangled SPARE models, SPARE-AD2 and SPARE-BA3, derived brain patterns that were more specific to the two types of brain changes. The correlation between the SPARE-BA Gap (SPARE-BA minus chronological age) and SPARE-AD was significantly reduced after the decoupling (r = 0.56–0.06). The correlation of disentangled SPARE-AD was non-inferior to amyloid- and tau-related measurements and to the number of APOE ε4 alleles but was lower to Alzheimer’s disease-related psychometric test scores, suggesting the contribution of advanced brain ageing to the latter. The disentangled SPARE-BA was consistently less correlated with Alzheimer’s disease-related clinical, molecular and genetic variables.

By employing conservative molecular diagnoses and introducing Alzheimer’s disease continuum cases to the SPARE-BA model training, we achieved more dissociable neuroanatomical biomarkers of typical brain ageing and Alzheimer’s disease.

Characterizing Heterogeneity in Neuroimaging, Cognition, Clinical Symptoms, and Genetics Among Patients With Late-Life Depression

Importance

Late-life depression (LLD) is characterized by considerable heterogeneity in clinical manifestation. Unraveling such heterogeneity might aid in elucidating etiological mechanisms and support precision and individualized medicine.

Objective

To cross-sectionally and longitudinally delineate disease-related heterogeneity in LLD associated with neuroanatomy, cognitive functioning, clinical symptoms, and genetic profiles.

Design, Setting, and Participants

The Imaging-Based Coordinate System for Aging and Neurodegenerative Diseases (iSTAGING) study is an international multicenter consortium investigating brain aging in pooled and harmonized data from 13 studies with more than 35 000 participants, including a subset of individuals with major depressive disorder. Multimodal data from a multicenter sample (N = 996), including neuroimaging, neurocognitive assessments, and genetics, were analyzed in this study. A semisupervised clustering method (heterogeneity through discriminative analysis) was applied to regional gray matter (GM) brain volumes to derive dimensional representations. Data were collected from July 2017 to July 2020 and analyzed from July 2020 to December 2021.

Main Outcomes and Measures

Two dimensions were identified to delineate LLD-associated heterogeneity in voxelwise GM maps, white matter (WM) fractional anisotropy, neurocognitive functioning, clinical phenotype, and genetics.

Results

A total of 501 participants with LLD (mean [SD] age, 67.39 [5.56] years; 332 women) and 495 healthy control individuals (mean [SD] age, 66.53 [5.16] years; 333 women) were included. Patients in dimension 1 demonstrated relatively preserved brain anatomy without WM disruptions relative to healthy control individuals. In contrast, patients in dimension 2 showed widespread brain atrophy and WM integrity disruptions, along with cognitive impairment and higher depression severity. Moreover, 1 de novo independent genetic variant (rs13120336; chromosome: 4, 186387714; minor allele, G) was significantly associated with dimension 1 (odds ratio, 2.35; SE, 0.15; P = 3.14 ×108) but not with dimension 2. The 2 dimensions demonstrated significant single-nucleotide variant-based heritability of 18% to 27% within the general population (N = 12 518 in UK Biobank). In a subset of individuals having longitudinal measurements, those in dimension 2 experienced a more rapid longitudinal change in GM and brain age (Cohen f2 = 0.03; P = .02) and were more likely to progress to Alzheimer disease (Cohen f2 = 0.03; P = .03) compared with those in dimension 1 (N = 1431 participants and 7224 scans from the Alzheimer's Disease Neuroimaging Initiative [ADNI], Baltimore Longitudinal Study of Aging [BLSA], and Biomarkers for Older Controls at Risk for Dementia [BIOCARD] data sets).

Conclusions and Relevance

This study characterized heterogeneity in LLD into 2 dimensions with distinct neuroanatomical, cognitive, clinical, and genetic profiles. This dimensional approach provides a potential mechanism for investigating the heterogeneity of LLD and the relevance of the latent dimensions to possible disease mechanisms, clinical outcomes, and responses to interventions.

EPCO-25. MULTI-OMICS DISEASE STRATIFICATION IN PATIENTS WITH IDH-WILDTYPE GLIOBLASTOMA: SYNERGISTIC VALUE OF CLINICAL MEASURES, CONVENTIONAL AND DEEP RADIOMICS, AND GENOMICS FOR PREDICTION OF OVERALL SURVIVAL

PURPOSE

Multi-omics data integration captures tumor characteristics at multiple scales [i.e., microscopic (genomics and epigenetics), macroscopic (radiomics), clinical manifestation], provides a more comprehensive assessment of patient’s risk, and facilitates personalized therapies. In this work, we investigated the synergistic value of such multiple data sources for risk stratification and prediction of overall survival in IDH-wildtype glioblastoma tumors.

METHODS

Quantitative conventional and deep radiomics were extracted from pre-operative multi-parametric structural MRI (T1, T1Gd, T2, T2-FLAIR) of 501 patients with newly diagnosed glioblastoma. 389/501 and 112/501 patients formed our discovery and replication cohorts, respectively. Conventional radiomics were extracted from CaPTk, and deep radiomics from a pre-trained VGG-19 model. Multivariate SVM classification was performed on the discovery cohort to stratify patients in high, medium, and low-risk groups, using recursive feature elimination and 5-fold cross-validation. This model was independently tested on the replication cohort, and a radiomic-based survival prediction index (SPIradiomics) was calculated for each patient. Multi-stage integration of omics data, i.e., clinical (age, gender, extent of resection (EOR)), SPIradiomics, epigenetics (MGMT promoter methylation), and genomics (27 clinically relevant gene mutations via next-generation sequencing (NGS)), was performed using multivariate Cox proportional hazards (Cox-PH) model for stratification of the risk in the replication cohort.

RESULTS

Cox-PH modeling resulted in a concordance index (c-index) of 0.65 (95% CI:0.6–0.7) for clinical data, 0.67 (95% CI:0.62–0.72) for clinical and epigenetics, 0.70 (95% CI:0.65–0.75) for clinical and radiomics, 0.72 (95% CI:0.68–0.77) for clinical, epigenetics, and radiomics, and 0.75 (95% CI:0.71 – 0.78) for the multi-omics combination of all data; highlighting the added value of each layer of information in prediction of the patient’s risk.

CONCLUSION

Our results reinforce the synergistic value of integrated diagnostic methods for improving risk assessment of patients with glioblastoma that may pave the path towards a more personalized treatment planning.

NIMG-58. CANONICAL CORRELATION ANALYSIS IN GLIOBLASTOMA REVEALS ASSOCIATIONS BETWEEN EXPRESSION OF RADIOMIC SIGNATURES AND GENOMICS

PURPOSE

Understanding the molecular underpinnings of imaging signatures of glioblastoma can provide insights into the biologic basis of tumor formation and progression as well as in vivo surrogate markers of molecular events driving the tumor’s phenotype. Through machine learning (ML), this study demonstrates that distinct imaging subtypes of glioblastoma are related to specific molecular alterations.

METHODS

Pre-operative multi-parametric MRI (T1, T2, T1CE, T2-FLAIR, DSC-MRI, DTI-MRI) of 669 IDH-wildtype subjects with glioblastoma were retrospectively collected and radiomic features, including descriptors of morphology, intensity, histogram, and texture, were extracted. Imaging subtypes were identified by a feature selection and clustering approach. Genomic markers, obtained using next generation sequencing (NGS) panel of 27 key glioblastoma genes, were available in 358/669 patients. Canonical correlation analysis (CCA) was conducted within each imaging subtype between the selected imaging features and genetic variables to seek maximum correlations between combinations of variables in imaging and genomic sets, and hence elucidate the molecular drivers of respective subtypes.

RESULTS

Three distinct imaging subtypes were identified by clustering on 50 selected features, representing characteristics of morphology, tumor neo-angiogenesis (DSC-derived features), and cellular density (DTI-derived features). These subtypes yielded differentiable overall survival based on Kaplan-Meier analysis. The canonical coefficients of each subtype revealed the distinction of the underlying genomic characteristics: one exhibited frequently mutated [ARID2, NTRK1], another subtype showed increased frequency of mutation in [ATRX, EGFR, PIK3R1], while the third was associated with all these genes and [NF1, PIK3CA, RB1], additionally.

CONCLUSION

We derived three distinct radiomic MRI subtypes for glioblastoma that highly correlate with the patients' survival and molecular genetic characteristics. Investigating the relationship between imaging and genomic information may enable identification of molecularly- and phenotypically-consistent tumor subtypes, which would offer non-invasive approaches for characterizing heterogeneity of glioblastoma, further facilitating patient stratification and treatment planning.

NIMG-73. CAPTURING GLIOBLASTOMA HETEROGENEITY USING IMAGING AND DEEP LEARNING: APPLICATION TO MGMT PROMOTER METHYLATION

PURPOSE

Intratumor heterogeneity is frequent in glioblastoma (GB), giving rise to the tumor’s resistance to standard therapies and, ultimately, poorer clinical outcomes. Yet heterogeneity is often not quantified when assessing the genomic or methylomic profile of a tumor, when a single tissue sample is analyzed. This study proposes a novel approach to non-invasively characterize heterogeneity across glioblastoma using deep learning analysis MRI scans, using MGMT promoter methylation (MGMTpm) as a test-case, and validates the imaging-derived heterogeneity maps with MGMTpm heterogeneity measured via multiple tissue samples.

METHODS

Multi-parametric MRI (mpMRI) scans (T1, T1-Gd, T2, T2-FLAIR) of 181 patients with newly diagnosed glioblastoma, who underwent surgical tumor resection and had MGMT methylation assessment results, were retrospectively collected. We trained a 5-fold cross-validated deep convolutional neural network with six convolutional layers for a discovery cohort of 137 patients by placing overlapping regional patches over the whole tumor on mpMRI scans to capture spatial heterogeneity of MGMTpm status in different regions within the tumor. Our approach effectively hypothesized that despite heterogeneity in the training examples, dominant imaging patterns would be captured by deep learning. Trained model was independently applied to an unseen replication cohort of 44 patients, with multiple tissue specimens chosen from different spatial regions within the tumor, allowing us to compare imaging- and tissue-based MGMTpm estimates.

RESULTS

Our model yielded AUC of 0.75 (95% CI: 0.65–0.79) for global MGMT status prediction, which reflected the heterogeneity in MGMTpm, but also that a dominant imaging pattern of MGMT methylation seemed to emerge. In methylated patients with multiple tissue samples, a significant Pearson's correlation coefficient of 0.64 (p&lt; 0.05) was found between imaging-based heterogeneity maps and MGMTpm heterogeneity.

CONCLUSION

A novel method based on mpMRI and deep neural networks yielded imaging-based heterogeneity maps that strongly associated with intratumor molecular heterogeneity in MGMT promoter methylated tumors.

NIMG-28. PROSPECTIVE HISTOPATHOLOGY-VALIDATED MACHINE LEARNING FOR DISTINGUISHING TRUE PROGRESSION FROM TREATMENT-RELATED CHANGES IN GLIOBLASTOMA PATIENTS

PURPOSE

Decision making about the best course of treatment for glioblastoma patients becomes challenging when a new enhancing lesion appears in the vicinity of the surgical bed on follow-up MRI (after maximal safe tumor resection and chemoradiation), raising concerns for tumor progression (TP). Literature indicates 30-50% of these new lesions describe primarily treatment-related changes (TRC). We hypothesize that quantitative analysis of specific and sensitive features extracted from multi-parametric MRI (mpMRI) via machine learning (ML) techniques may yield non-invasive imaging signatures that distinguish TP from TRC and facilitate better treatment personalization.

METHODS

We have generated an ML model on a retrospective cohort of 58 subjects, and prospectively evaluated on an independent cohort of 58 previously unseen patients who underwent second resection for suspicious recurrence and had availability of advanced mpMRI (T1, T1-Gd, T2, T2-FLAIR, DTI, DSC). The features selected by our retrospective model, representing principal components analysis of intensity distributions, morphological, statistical, and texture descriptors, were extracted from the mpMRI of the prospective cohort. Integration of these features revealed signatures distinguishing between TP, mixed response, and TRC. Independently, a board-certified neuropathologist evaluated the resected tissue by blindly classifying it in the above three categories, based on mitotic figures, pseudopalisading necrosis, geographic necrosis, dystrophic calcification, vascular changes, and Ki67.

RESULTS

Tissues classified as TRC by the neuropathologist were associated with imaging phenotypes of lower angiogenesis (DSC-derived features), lower cellularity (DTI-derived features), and higher water concentration (T2, T2-FLAIR features). Our ML model characterized TP with 78% accuracy (sensitivity:86%, specificity:70%, AUC:0.80 (95%CI, 0.68-0.92)) and TRC with 81% accuracy (sensitivity:80%, specificity:81%, AUC:0.87 (95%CI, 0.72-1.00)).

CONCLUSION

Our proposed ML model reveals distinct non-invasive markers of TP and TRC, directly associated with histopathological changes in prospective glioblastoma patients. Reliable stratification of TP and TRC entities may help to noninvasively determine whether the course of treatment should change.

NIMG-52. RADIOGENOMICS SIGNATURES IN KEY DRIVER GENES IN GLIOBLASTOMA EVALUATED WITH AND WITHOUT THE PRESENCE OF CO-OCCURRING MUTATIONS

PURPOSE

Glioblastomas display significant heterogeneity on the molecular level, typically harboring several co-occurring mutations, which likely contributes to failure of molecularly targeted therapeutic approaches. Radiogenomics has emerged as a promising tool for in vivo characterization of this heterogeneity. We derive radiogenomic signatures of four mutations via machine learning (ML) analysis of multiparametric MRI (mpMRI) and evaluate them in the presence and absence of other co-occurring mutations.

METHODS

We identified a retrospective cohort of 359 IDH-wildtype glioblastoma patients, with available pre-operative mpMRI (T1, T1Gd, T2, T2-FLAIR) scans and targeted next generation sequencing (NGS) data. Radiomic features, including morphologic, histogram, texture, and Gabor wavelet descriptors, were extracted from the mpMRI. Multivariate predictive models were trained using cross-validated SVM with LASSO feature selection to predict mutation status in key driver genes, EGFR, PTEN, TP53, and NF1. ML models and spatial population atlases of genetic mutations were generated for stratification of the tumors (1) with co-occurring mutations versus wildtypes, (2) with exclusive mutations in each driver gene versus the tumors without any mutations in the pathways associated with these genes.

RESULTS

ML models yielded AUCs of 0.75 (95%CI:0.62-0.88) / 0.87 (95%CI:0.70-1) for co-occurring / exclusive EGFR mutations, 0.69 (95%CI:0.58-0.80) / 0.80 (95%CI:0.61-0.99) for co-occurring / exclusive PTEN mutations, and 0.77 (95%CI:0.65-0.88) / 0.86 (95%CI:0.69-1) for co-occurring / exclusive TP53 cases. Spatial atlases revealed a predisposition of left temporal lobe for NF1 and right frontotemporal region for TP53 in mutually exclusive tumors, which was not observed in the co-occurring mutation atlases.

CONCLUSION

Our results suggest the presence of distinct radiogenomic signatures of several glioblastoma mutations, which become even more pronounced when respective mutations do not co-occur with other mutations. These in vivo signatures can contribute to pre-operative stratification of patients for molecular targeted therapies, and potentially longitudinal monitoring of mutational changes during treatment.

Quantification of tumor microenvironment acidity in glioblastoma using principal component analysis of dynamic susceptibility contrast enhanced MR imaging

Glioblastoma (GBM) has high metabolic demands, which can lead to acidification of the tumor microenvironment. We hypothesize that a machine learning model built on temporal principal component analysis (PCA) of dynamic susceptibility contrast-enhanced (DSC) perfusion MRI can be used to estimate tumor acidity in GBM, as estimated by pH-sensitive amine chemical exchange saturation transfer echo-planar imaging (CEST-EPI). We analyzed 78 MRI scans in 32 treatment naïve and post-treatment GBM patients. All patients were imaged with DSC-MRI, and pH-weighting that was quantified from CEST-EPI estimation of the magnetization transfer ratio asymmetry (MTRasym) at 3 ppm. Enhancing tumor (ET), non-enhancing core (NC), and peritumoral T2 hyperintensity (namely, edema, ED) were used to extract principal components (PCs) and to build support vector machines regression (SVR) models to predict MTRasym values using PCs. Our predicted map correlated with MTRasym values with Spearman's r equal to 0.66, 0.47, 0.67, 0.71, in NC, ET, ED, and overall, respectively (p < 0.006). The results of this study demonstrates that PCA analysis of DSC imaging data can provide information about tumor pH in GBM patients, with the strongest association within the peritumoral regions.

NIMG-22. INTEGRATION OF A RADIOMIC SIGNATURE, CLINICAL VARIABLES AND PLASMA CELL-FREE DNA IN ADULT PATIENTS WITH NEWLY DIAGNOSED GLIOBLASTOMA PREDICTS PATIENT SURVIVAL AND IMPROVES DISEASE STRATIFICATION

PURPOSE

We have previously demonstrated the potential role of liquid biopsy, specifically plasma cell-free DNA (cfDNA), as a non-invasive biomarker for prognostication in patients with glioblastoma. In separate prior studies, we have also developed MRI-based radiomic signatures to predict survival outcomes in glioblastoma. In this study, for the first time, we evaluated the potential of combining radiomic signatures, epidemiological and clinical variables, and plasma cfDNA quantification for upfront prediction of overall survival (OS) in patients with newly diagnosed glioblastoma.

METHODS

Quantitative radiomic features were extracted from multiparametric MRI (T1, T1Gd, T2, T2-FLAIR) scans of a discovery cohort of 505 and an independent replication cohort of 50 IDH-wildtype glioblastoma patients. For the independent replication cohort, pre-surgical plasma cfDNA was extracted and quantified. In the first stage, a radiomic signature was created for stratification of patients into categories of short (OS ≤ 6 months) and long (OS ≥ 18 months) survivors using a cross-validated XGBoost method based on the discovery cohort, which was tested independently on the replication cohort. In the second stage, the radiomic signature and clinical variables were integrated to build a second-stage signature using a cross-validated support vector machine (SVM) classifier to stratify the patients into short and long survivor categories. In the third stage, the value of the second-stage signature integrated with cfDNA concentration was assessed through a cross-validated SVM regression method.

RESULTS

The combination of radiomic, clinical, and cfDNA variables resulted in the best overall predictive accuracy, with Pearson’s correlation coefficient of 0.59 (p&lt; 0.0001) between actual and predicted OS.

CONCLUSION

In this study, we evaluated the value of combining plasma cfDNA, radiomic, and clinical variables for predicting OS, and showed that it could act as an effective non-invasive prognostic and patient stratification tool in patients with newly diagnosed glioblastoma.

NIMG-40. RADIOGENOMIC SIGNATURES OF DRIVER GENES IN NEWLY DIAGNOSED GLIOBLASTOMA PATIENTS BASED ON PRE-OPERATIVE MULTI-PARAMETRIC MRI

PURPOSE

Non-invasive and quantitative biomarkers of somatic mutations derived from multi-parametric MRI (MP-MRI) could potentially help in predicting the response of patients to therapy, leading to development of targeted and personalized treatments. In this study, we developed radiogenomic signatures of multiple driver genes using artificial intelligence (AI) methods.

METHODS

In this study, 2740 radiomic features, including shape and volumetric measures computed for different tumorous regions, and characteristics derived from histograms and gray-level co-occurrence matrix (GLCM), were extracted from pre-operative MP-MRI (T1, T1Gd, T2, T2-FLAIR, DTI, and DSC-MRI) scans of 161 patients with newly diagnosed glioblastoma. The tumor samples, collected surgically from these patients, were sequenced using an in-house targeted next generation sequencing (NGS) panel of genes. We constructed quantitative imaging signatures of somatic mutations in several genes from 161 IDH-wildtype glioblastoma patients, including ATRX, FGFR2, EGFR, MET, NF1, PDGFRA, PIK3CA, PTEN, RB1, TP53, using cross-validated SVM classifiers.

RESULTS

The cross-validated classification performance for each signature was assessed by area under the receiver operating characteristic (ROC) curve (AUC), indicating the following results: PTEN (n = 69, AUC = 0.64), EGFR (n = 52, AUC = 0.72), TP53 (n = 51, AUC = 0.67), NF1 (n = 33, AUC = 0.74), ATRX (n = 22; AUC = 0.74), FGFR2 (n = 6, AUC = 0.82), MET (n = 26, AUC = 0.77), PDGFRA (n = 14, AUC = 0.82), PIK3CA (n = 14, AUC = 0.78), RB1 (n = 14, AUC = 0.81).

CONCLUSION

Using multi-parametric MRI, we developed quantitative non-invasive in vivo signatures with the potential for pre-operative assessment of a glioblastoma’s molecular characteristics. These non-invasive radiogenomic biomarkers may be useful for understanding the molecular composition of a glioblastoma prior to surgical resection, thus enabling earlier selection of patients for targeted therapy trials and possible neoadjuvant treatment.

The Brain Chart of Aging: Machine-learning analytics reveals links between brain aging, white matter disease, amyloid burden, and cognition in the iSTAGING consortium of 10,216 harmonized MR scans

INTRODUCTION

Relationships between brain atrophy patterns of typical aging and Alzheimer's disease (AD), white matter disease, cognition, and AD neuropathology were investigated via machine learning in a large harmonized magnetic resonance imaging database (11 studies; 10,216 subjects).

METHODS

Three brain signatures were calculated: Brain-age, AD-like neurodegeneration, and white matter hyperintensities (WMHs). Brain Charts measured and displayed the relationships of these signatures to cognition and molecular biomarkers of AD.

RESULTS

WMHs were associated with advanced brain aging, AD-like atrophy, poorer cognition, and AD neuropathology in mild cognitive impairment (MCI)/AD and cognitively normal (CN) subjects. High WMH volume was associated with brain aging and cognitive decline occurring in an ≈10-year period in CN subjects. WMHs were associated with doubling the likelihood of amyloid beta (Aβ) positivity after age 65. Brain aging, AD-like atrophy, and WMHs were better predictors of cognition than chronological age in MCI/AD.

DISCUSSION

A Brain Chart quantifying brain-aging trajectories was established, enabling the systematic evaluation of individuals' brain-aging patterns relative to this large consortium.

Histopathology-validated machine learning radiographic biomarker for noninvasive discrimination between true progression and pseudo-progression in glioblastoma

BACKGROUND

Imaging of glioblastoma patients after maximal safe resection and chemoradiation commonly demonstrates new enhancements that raise concerns about tumor progression. However, in 30% to 50% of patients, these enhancements primarily represent the effects of treatment, or pseudo-progression (PsP). We hypothesize that quantitative machine learning analysis of clinically acquired multiparametric magnetic resonance imaging (mpMRI) can identify subvisual imaging characteristics to provide robust, noninvasive imaging signatures that can distinguish true progression (TP) from PsP.

METHODS

We evaluated independent discovery (n = 40) and replication (n = 23) cohorts of glioblastoma patients who underwent second resection due to progressive radiographic changes suspicious for recurrence. Deep learning and conventional feature extraction methods were used to extract quantitative characteristics from the mpMRI scans. Multivariate analysis of these features revealed radiophenotypic signatures distinguishing among TP, PsP, and mixed response that compared with similar categories blindly defined by board-certified neuropathologists. Additionally, interinstitutional validation was performed on 20 new patients.

RESULTS

Patients who demonstrate TP on neuropathology are significantly different (P < .0001) from those with PsP, showing imaging features reflecting higher angiogenesis, higher cellularity, and lower water concentration. The accuracy of the proposed signature in leave-one-out cross-validation was 87% for predicting PsP (area under the curve [AUC], 0.92) and 84% for predicting TP (AUC, 0.83), whereas in the discovery/replication cohort, the accuracy was 87% for predicting PsP (AUC, 0.84) and 78% for TP (AUC, 0.80). The accuracy in the interinstitutional cohort was 75% (AUC, 0.80).

CONCLUSION

Quantitative mpMRI analysis via machine learning reveals distinctive noninvasive signatures of TP versus PsP after treatment of glioblastoma. Integration of the proposed method into clinical studies can be performed using the freely available Cancer Imaging Phenomics Toolkit.

Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan

As medical imaging enters its information era and presents rapidly increasing needs for big data analytics, robust pooling and harmonization of imaging data across diverse cohorts with varying acquisition protocols have become critical. We describe a comprehensive effort that merges and harmonizes a large-scale dataset of 10,477 structural brain MRI scans from participants without a known neurological or psychiatric disorder from 18 different studies that represent geographic diversity. We use this dataset and multi-atlas-based image processing methods to obtain a hierarchical partition of the brain from larger anatomical regions to individual cortical and deep structures and derive age trends of brain structure through the lifespan (3-96 years old). Critically, we present and validate a methodology for harmonizing this pooled dataset in the presence of nonlinear age trends. We provide a web-based visualization interface to generate and present the resulting age trends, enabling future studies of brain structure to compare their data with this reference of brain development and aging, and to examine deviations from ranges, potentially related to disease.

NIMG-35. QUANTITATIVE ESTIMATION OF PROGRESSION-FREE SURVIVAL BASED ON RADIOMICS ANALYSIS OF PREOPERATIVE MULTI-PARAMETRIC MRI IN PATIENTS WITH GLIOBLASTOMA

PURPOSE

Glioblastoma managed with maximal resection and adjuvant chemoradiation has large heterogeneity at the time of tumor recurrence (progression-free survival, PFS). Upfront identification of patients with shorter than median PFS, may facilitate better personalization of treatment, such as patient stratification into clinical trials for treatment intensification and/or supportive care. The aim of the current study was to build a radiomics model that can reliably predict PFS in patients with glioblastoma at the time of initial diagnosis. EXPERIMENTAL DESIGN: A total of 66 de novo glioblastoma patients in a single institution who underwent gross total resection followed by standard adjuvant chemoradiation, were found to have pathologically confirmed recurrence. Patients who had sufficient follow-up MRI to reliably determine PFS and available pre-surgical multi-parametric MRI (mpMRI) (T1, T1Gd, T2, T2-FLAIR, DTI, DSC-MR images) were included. The patients were stratified into two classes of “short PFS (≤7 months, n=35)” and “long PFS (≥9 months, n=19)”. An extensive set of features was extracted from preoperative mpMRI scans, including shape and volumetric measures computed for different tumorous regions, and characteristics derived from histograms and gray-level co-occurrence matrix (GLCM). Predictive modeling for discrimination of patients with short from those with long PFS was performed using a 10-fold cross-validated support vector machine classifier.

RESULTS

An accuracy of 79.6% was achieved for prediction of short and long PFS (sensitivity=83%, specificity=74%, AUC=0.86 [95%CI: 0.76–0.96]; short-PSF was considered the positive class). The patients with shorter PFS showed markedly higher cell density in the nonenhancing tumor (DTI), higher neo-angiogenesis in the enhancing tumor (DSC-MRI, T1Gd), lower water concentration in the enhancing tumor region (T2), and higher maximum enhancement (calculated by the subtraction of T1 from T1Gd) in the edema region.

CONCLUSION

Quantitative assessment of preoperative mpMRI provides accurate prediction of PFS for patients with glioblastoma, which can aid in personalized treatment planning.

TMOD-40. IN VIVO EVALUATION OF O6-METHYLGUANINE-DNA-METHYLTRANSFERASE (MGMT) PROMOTER METHYLATION STATUS FOR DE NOVO GLIOBLASTOMA PATIENTS USING DEEP LEARNING FEATURES

BACKGROUND

High expression of O6-methylguanine-DNA methyltransferase (MGMT) in glioblastoma is associated with resistance to temozolomide, as tumor cells lacking MGMT activity are significantly more sensitive to the cytotoxic effects of temozolomide. The MGMT promoter methylation status (MGMTpms) is typically determined as MGMT-methylated or MGMT-unmethylated. Some single-center studies have reported results ranging from 70–95% detection rates using MRI. We aim to further validate these findings using a multi-institutional data set. We hypothesize that transfer learning based features when integrated via machine learning may lead to non-invasive determination of MGMTpms.

METHODS

A total of 270 patients were included across the 3 institutions (Hospital of the University of Pennsylvania (HUP), Jefferson University Hospital (JUH); the TCIA). JUH and TCIA datasets comprised conventional modalities (T1,T2,T2-FLAIR,T1-Gd), whereas HUP dataset had additional modalities (DSC,DTI) as well. We used transfer learning and adapted a convolutional neural network (CNN) model pre-trained on 1.2 million 3-channel images of the ImageNet to extract deep learning features from the given images. A support vector machine multivariately integrated these features towards a non-invasive marker of MGMTpms.

RESULTS

The cross-validated accuracy of our MGMT marker in classifying the mutation status in individual patients was 86.95%, 81.56%, and 82.43%, respectively, in HUP, JUH, and TCIA. Our marker revealed MGMT-methylated tumors with lower neovascularization and cell density, when compared with MGMT-unmethylated tumors. MGMT-unmethylated tumors were found to be more lateralized to the right hemisphere, when compared with MGMT-methylated tumors.

CONCLUSION

Our findings suggest that transfer learning features when integrated via machine learning allow robust prediction of MGMTpms on mpMRI acquired within multiple institutions. The proposed non-invasive MGMT marker may contribute to (i) MGMTpms determination for patients with inadequate tissue/inoperable tumors, (ii) stratification of patients into clinical trials, (iii) patient selection for targeted therapy, and (iv) personalized treatment planning.