Data-driven modelling of neurodegenerative disease progression: thinking outside the black box.

Data-driven disease progression models are an emerging set of computational tools that reconstruct disease timelines for long-term chronic diseases, providing unique insights into disease processes and their underlying mechanisms. Such methods combine a priori human knowledge and assumptions with large-scale data processing and parameter estimation to infer long-term disease trajectories from short-term data. In contrast to 'black box' machine learning tools, data-driven disease progression models typically require fewer data and are inherently interpretable, thereby aiding disease understanding in addition to enabling classification, prediction and stratification. In this Review, we place the current landscape of data-driven disease progression models in a general framework and discuss their enhanced utility for constructing a disease timeline compared with wider machine learning tools that construct static disease profiles. We review the insights they have enabled across multiple neurodegenerative diseases, notably Alzheimer disease, for applications such as determining temporal trajectories of disease biomarkers, testing hypotheses about disease mechanisms and uncovering disease subtypes. We outline key areas for technological development and translation to a broader range of neuroscience and non-neuroscience applications. Finally, we discuss potential pathways and barriers to integrating disease progression models into clinical practice and trial settings.

A foundation model for generalizable disease detection from retinal images.

Medical artificial intelligence (AI) offers great potential for recognizing signs of health conditions in retinal images and expediting the diagnosis of eye diseases and systemic disorders1. However, the development of AI models requires substantial annotation and models are usually task-specific with limited generalizability to different clinical applications2. Here, we present RETFound, a foundation model for retinal images that learns generalizable representations from unlabelled retinal images and provides a basis for label-efficient model adaptation in several applications. Specifically, RETFound is trained on 1.6 million unlabelled retinal images by means of self-supervised learning and then adapted to disease detection tasks with explicit labels. We show that adapted RETFound consistently outperforms several comparison models in the diagnosis and prognosis of sight-threatening eye diseases, as well as incident prediction of complex systemic disorders such as heart failure and myocardial infarction with fewer labelled data. RETFound provides a generalizable solution to improve model performance and alleviate the annotation workload of experts to enable broad clinical AI applications from retinal imaging.

Towards cascading genetic risk in Alzheimer's disease.

Alzheimer's disease typically progresses in stages, which have been defined by the presence of disease-specific biomarkers: Amyloid (A), Tau (T) and neurodegeneration (N). This progression of biomarkers has been condensed into the ATN framework, where each of the biomarkers can be either positive (+) or negative (-). Over the past decades genome wide association studies have implicated about 90 different loci involved with the development of late onset Alzheimer's disease. Here we investigate whether genetic risk for Alzheimer's disease contributes equally to the progression in different disease stages or whether it exhibits a stage-dependent effect. Amyloid (A) and tau (T) status was defined using a combination of available PET and CSF biomarkers in the Alzheimer's Disease Neuroimaging Initiative cohort. In 312 participants with biomarker-confirmed A-T- status, we employed Cox proportional hazards models to estimate the contribution of APOE and polygenic risk scores (beyond APOE) to convert to A+T- status (65 conversions). Furthermore, we repeated the analysis in 290 participants with A+T- status and investigated the genetic contribution to conversion to A+T+ (45 conversions). Both survival analyses were adjusted for age, sex, and years of education. For progression from A-T- to A+T-, APOE-e4 burden showed significant effect (HR=2.88; 95% CI: 1.70-4.89; P<0.001), while polygenic risk did not (HR=1.09; 95% CI: 0.84-1.42; P=0.53). Conversely, for the transition from A+T- to A+T+, the APOE-e4 burden contribution was reduced (HR=1.62 95% CI: 1.05-2.51; P=0.031), while the polygenic risk showed an increased contribution (HR=1.73; 95% CI:1.27-2.36; P<0.001). The marginal APOE effect was driven by e4 homozygotes (HR=2.58; 95% CI: 1.05-6.35; P=0.039) as opposed to e4 heterozygotes (HR=1.74; 95% CI: 0.87-3.49; P=0.12). The genetic risk for late-onset Alzheimer's disease unfolds in a disease stage-dependent fashion. A better understanding of the interplay between disease stage and genetic risk can lead to a more mechanistic understanding of transition between ATN stages, a better understanding of the molecular processes leading to Alzheimer's disease as well as opening therapeutic windows for targeted interventions.

Mortality surrogates in combined pulmonary fibrosis and emphysema.

BACKGROUND: Idiopathic pulmonary fibrosis (IPF) with coexistent emphysema, termed combined pulmonary fibrosis and emphysema (CPFE) may associate with reduced forced vital capacity (FVC) declines compared to non-CPFE IPF patients. We examined associations between mortality and functional measures of disease progression in two IPF cohorts. METHODS: Visual emphysema presence (>0% emphysema) scored on computed tomography identified CPFE patients (CPFE/non-CPFE: derivation cohort n=317/n=183, replication cohort n=358/n=152), who were subgrouped using 10% or 15% visual emphysema thresholds, and an unsupervised machine-learning model considering emphysema and interstitial lung disease extents. Baseline characteristics, 1-year relative FVC and diffusing capacity of the lung for carbon monoxide (D LCO) decline (linear mixed-effects models), and their associations with mortality (multivariable Cox regression models) were compared across non-CPFE and CPFE subgroups. RESULTS: In both IPF cohorts, CPFE patients with ≥10% emphysema had a greater smoking history and lower baseline D LCO compared to CPFE patients with <10% emphysema. Using multivariable Cox regression analyses in patients with ≥10% emphysema, 1-year D LCO decline showed stronger mortality associations than 1-year FVC decline. Results were maintained in patients suitable for therapeutic IPF trials and in subjects subgrouped by ≥15% emphysema and using unsupervised machine learning. Importantly, the unsupervised machine-learning approach identified CPFE patients in whom FVC decline did not associate strongly with mortality. In non-CPFE IPF patients, 1-year FVC declines ≥5% and ≥10% showed strong mortality associations. CONCLUSION: When assessing disease progression in IPF, D LCO decline should be considered in patients with ≥10% emphysema and a ≥5% 1-year relative FVC decline threshold considered in non-CPFE IPF patients.

A data-driven study of Alzheimer's disease related amyloid and tau pathology progression.

Amyloid-ß is thought to facilitate the spread of tau throughout the neocortex in Alzheimer's disease, though how this occurs is not well understood. This is because of the spatial discordance between amyloid-ß, which accumulates in the neocortex, and tau, which accumulates in the medial temporal lobe during ageing. There is evidence that in some cases amyloid-ß-independent tau spreads beyond the medial temporal lobe where it may interact with neocortical amyloid-ß. This suggests that there may be multiple distinct spatiotemporal subtypes of Alzheimer's-related protein aggregation, with potentially different demographic and genetic risk profiles. We investigated this hypothesis, applying data-driven disease progression subtyping models to post-mortem neuropathology and in vivo PET-based measures from two large observational studies: the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Religious Orders Study and Rush Memory and Aging Project (ROSMAP). We consistently identified 'amyloid-first' and 'tau-first' subtypes using cross-sectional information from both studies. In the amyloid-first subtype, extensive neocortical amyloid-ß precedes the spread of tau beyond the medial temporal lobe, while in the tau-first subtype, mild tau accumulates in medial temporal and neocortical areas prior to interacting with amyloid-ß. As expected, we found a higher prevalence of the amyloid-first subtype among apolipoprotein E (APOE) ε4 allele carriers while the tau-first subtype was more common among APOE ε4 non-carriers. Within tau-first APOE ε4 carriers, we found an increased rate of amyloid-ß accumulation (via longitudinal amyloid PET), suggesting that this rare group may belong within the Alzheimer's disease continuum. We also found that tau-first APOE ε4 carriers had several fewer years of education than other groups, suggesting a role for modifiable risk factors in facilitating amyloid-ß-independent tau. Tau-first APOE ε4 non-carriers, in contrast, recapitulated many of the features of primary age-related tauopathy. The rate of longitudinal amyloid-ß and tau accumulation (both measured via PET) within this group did not differ from normal ageing, supporting the distinction of primary age-related tauopathy from Alzheimer's disease. We also found reduced longitudinal subtype consistency within tau-first APOE ε4 non-carriers, suggesting additional heterogeneity within this group. Our findings support the idea that amyloid-ß and tau may begin as independent processes in spatially disconnected regions, with widespread neocortical tau resulting from the local interaction of amyloid-ß and tau. The site of this interaction may be subtype-dependent: medial temporal lobe in amyloid-first, neocortex in tau-first. These insights into the dynamics of amyloid-ß and tau may inform research and clinical trials that target these pathologies.

Identification of different MRI atrophy progression trajectories in epilepsy by subtype and stage inference.

Artificial intelligence (AI)-based tools are widely employed, but their use for diagnosis and prognosis of neurological disorders is still evolving. Here we analyse a cross-sectional multicentre structural MRI dataset of 696 people with epilepsy and 118 control subjects. We use an innovative machine-learning algorithm, Subtype and Stage Inference, to develop a novel data-driven disease taxonomy, whereby epilepsy subtypes correspond to distinct patterns of spatiotemporal progression of brain atrophy.In a discovery cohort of 814 individuals, we identify two subtypes common to focal and idiopathic generalized epilepsies, characterized by progression of grey matter atrophy driven by the cortex or the basal ganglia. A third subtype, only detected in focal epilepsies, was characterized by hippocampal atrophy. We corroborate external validity via an independent cohort of 254 people and confirm that the basal ganglia subtype is associated with the most severe epilepsy.Our findings suggest fundamental processes underlying the progression of epilepsy-related brain atrophy. We deliver a novel MRI- and AI-guided epilepsy taxonomy, which could be used for individualized prognostics and targeted therapeutics.

Transferability of Alzheimer's disease progression subtypes to an independent population cohort.

In the past, methods to subtype or biotype patients using brain imaging data have been developed. However, it is unclear whether and how these trained machine learning models can be successfully applied to population cohorts to study the genetic and lifestyle factors underpinning these subtypes. This work, using the Subtype and Stage Inference (SuStaIn) algorithm, examines the generalisability of data-driven Alzheimer's disease (AD) progression models. We first compared SuStaIn models trained separately on Alzheimer's disease neuroimaging initiative (ADNI) data and an AD-at-risk population constructed from the UK Biobank dataset. We further applied data harmonization techniques to remove cohort effects. Next, we built SuStaIn models on the harmonized datasets, which were then used to subtype and stage subjects in the other harmonized dataset. The first key finding is that three consistent atrophy subtypes were found in both datasets, which match the previously identified subtype progression patterns in AD: 'typical', 'cortical' and 'subcortical'. Next, the subtype agreement was further supported by high consistency in individuals' subtypes and stage assignment based on the different models: more than 92% of the subjects, with reliable subtype assignment in both ADNI and UK Biobank dataset, were assigned to an identical subtype under the model built on the different datasets. The successful transferability of AD atrophy progression subtypes across cohorts capturing different phases of disease development enabled further investigations of associations between AD atrophy subtypes and risk factors. Our study showed that (1) the average age is highest in the typical subtype and lowest in the subcortical subtype; (2) the typical subtype is associated with statistically more-AD-like cerebrospinal fluid biomarkers values in comparison to the other two subtypes; and (3) in comparison to the subcortical subtype, the cortical subtype subjects are more likely to associate with prescription of cholesterol and high blood pressure medications. In summary, we presented cross-cohort consistent recovery of AD atrophy subtypes, showing how the same subtypes arise even in cohorts capturing substantially different disease phases. Our study opened opportunities for future detailed investigations of atrophy subtypes with a broad range of early risk factors, which will potentially lead to a better understanding of the disease aetiology and the role of lifestyle and behaviour on AD.

Accurate Bayesian segmentation of thalamic nuclei using diffusion MRI and an improved histological atlas.

The human thalamus is a highly connected brain structure, which is key for the control of numerous functions and is involved in several neurological disorders. Recently, neuroimaging studies have increasingly focused on the volume and connectivity of the specific nuclei comprising this structure, rather than looking at the thalamus as a whole. However, accurate identification of cytoarchitectonically designed histological nuclei on standard in vivo structural MRI is hampered by the lack of image contrast that can be used to distinguish nuclei from each other and from surrounding white matter tracts. While diffusion MRI may offer such contrast, it has lower resolution and lacks some boundaries visible in structural imaging. In this work, we present a Bayesian segmentation algorithm for the thalamus. This algorithm combines prior information from a probabilistic atlas with likelihood models for both structural and diffusion MRI, allowing segmentation of 25 thalamic labels per hemisphere informed by both modalities. We present an improved probabilistic atlas, incorporating thalamic nuclei identified from histology and 45 white matter tracts surrounding the thalamus identified in ultra-high gradient strength diffusion imaging. We present a family of likelihood models for diffusion tensor imaging, ensuring compatibility with the vast majority of neuroimaging datasets that include diffusion MRI data. The use of these diffusion likelihood models greatly improves identification of nuclear groups versus segmentation based solely on structural MRI. Dice comparison of 5 manually identifiable groups of nuclei to ground truth segmentations show improvements of up to 10 percentage points. Additionally, our chosen model shows a high degree of reliability, with median test-retest Dice scores above 0.85 for four out of five nuclei groups, whilst also offering improved detection of differential thalamic involvement in Alzheimer's disease (AUROC 81.98%). The probabilistic atlas and segmentation tool will be made publicly available as part of the neuroimaging package FreeSurfer (https://freesurfer.net/fswiki/ThalamicNucleiDTI).

Italian, European, and international neuroinformatics efforts: An overview.

Neuroinformatics is a research field that focusses on software tools capable of identifying, analysing, modelling, organising and sharing multiscale neuroscience data. Neuroinformatics has exploded in the last two decades with the emergence of the Big Data phenomenon, characterised by the so-called 3Vs (volume, velocity and variety), which provided neuroscientists with an improved ability to acquire and process data faster and more cheaply thanks to technical improvements in clinical, genomic and radiological technologies. This situation has led to a 'data deluge', as neuroscientists can routinely collect more study data in a few days than they could in a year just a decade ago. To address this phenomenon, several neuroimaging-focussed neuroinformatics platforms have emerged, funded by national or transnational agencies, with the following goals: (i) development of tools for archiving and organising analytical data (XNAT, REDCap and LabKey); (ii) development of data-driven models evolving from reductionist approaches to multidimensional models (RIN, IVN, HBD, EuroPOND, E-DADS and GAAIN BRAIN); and (iii) development of e-infrastructures to provide sufficient computational power and storage resources (neuGRID, HBP-EBRAINS, LONI and CONP). Although the scenario is still fragmented, there are technological and economical attempts at both national and international levels to introduce high standards for open and Findable, Accessible, Interoperable and Reusable (FAIR) neuroscience worldwide.

A framework for advancing sustainable magnetic resonance imaging access in Africa.

Magnetic resonance imaging (MRI) technology has profoundly transformed current healthcare systems globally, owing to advances in hardware and software research innovations. Despite these advances, MRI remains largely inaccessible to clinicians, patients, and researchers in low-resource areas, such as Africa. The rapidly growing burden of noncommunicable diseases in Africa underscores the importance of improving access to MRI equipment as well as training and research opportunities on the continent. The Consortium for Advancement of MRI Education and Research in Africa (CAMERA) is a network of African biomedical imaging experts and global partners, implementing novel strategies to advance MRI access and research in Africa. Upon its inception in 2019, CAMERA sets out to identify challenges to MRI usage and provide a framework for addressing MRI needs in the region. To this end, CAMERA conducted a needs assessment survey (NAS) and a series of symposia at international MRI society meetings over a 2-year period. The 68-question NAS was distributed to MRI users in Africa and was completed by 157 clinicians and scientists from across Sub-Saharan Africa (SSA). On average, the number of MRI scanners per million people remained at less than one, of which 39% were obsolete low-field systems but still in use to meet daily clinical needs. The feasibility of coupling stable energy supplies from various sources has contributed to the growing number of higher-field (1.5 T) MRI scanners in the region. However, these systems are underutilized, with only 8% of facilities reporting clinical scans of 15 or more patients per day, per scanner. The most frequently reported MRI scans were neurological and musculoskeletal. The CAMERA NAS combined with the World Health Organization and International Atomic Energy Agency data provides the most up-to-date data on MRI density in Africa and offers a unique insight into Africa's MRI needs. Reported gaps in training, maintenance, and research capacity indicate ongoing challenges in providing sustainable high-value MRI access in SSA. Findings from the NAS and focused discussions at international MRI society meetings provided the basis for the framework presented here for advancing MRI capacity in SSA. While these findings pertain to SSA, the framework provides a model for advancing imaging needs in other low-resource settings.

Automated airway quantification associates with mortality in idiopathic pulmonary fibrosis.

OBJECTIVES: The study examined whether quantified airway metrics associate with mortality in idiopathic pulmonary fibrosis (IPF). METHODS: In an observational cohort study (n = 90) of IPF patients from Ege University Hospital, an airway analysis tool AirQuant calculated median airway intersegmental tapering and segmental tortuosity across the 2nd to 6th airway generations. Intersegmental tapering measures the difference in median diameter between adjacent airway segments. Tortuosity evaluates the ratio of measured segmental length against direct end-to-end segmental length. Univariable linear regression analyses examined relationships between AirQuant variables, clinical variables, and lung function tests. Univariable and multivariable Cox proportional hazards models estimated mortality risk with the latter adjusted for patient age, gender, smoking status, antifibrotic use, CT usual interstitial pneumonia (UIP) pattern, and either forced vital capacity (FVC) or diffusion capacity of carbon monoxide (DLco) if obtained within 3 months of the CT. RESULTS: No significant collinearity existed between AirQuant variables and clinical or functional variables. On univariable Cox analyses, male gender, smoking history, no antifibrotic use, reduced DLco, reduced intersegmental tapering, and increased segmental tortuosity associated with increased risk of death. On multivariable Cox analyses (adjusted using FVC), intersegmental tapering (hazard ratio (HR) = 0.75, 95% CI = 0.66-0.85, p < 0.001) and segmental tortuosity (HR = 1.74, 95% CI = 1.22-2.47, p = 0.002) independently associated with mortality. Results were maintained with adjustment using DLco. CONCLUSIONS: AirQuant generated measures of intersegmental tapering and segmental tortuosity independently associate with mortality in IPF patients. Abnormalities in proximal airway generations, which are not typically considered to be abnormal in IPF, have prognostic value. CLINICAL RELEVANCE STATEMENT: Quantitative measurements of intersegmental tapering and segmental tortuosity, in proximal (second to sixth) generation airway segments, independently associate with mortality in IPF. Automated airway analysis can estimate disease severity, which in IPF is not restricted to the distal airway tree. KEY POINTS: • AirQuant generates measures of intersegmental tapering and segmental tortuosity. • Automated airway quantification associates with mortality in IPF independent of established measures of disease severity. • Automated airway analysis could be used to refine patient selection for therapeutic trials in IPF.

Learning from multiple annotators for medical image segmentation.

Supervised machine learning methods have been widely developed for segmentation tasks in recent years. However, the quality of labels has high impact on the predictive performance of these algorithms. This issue is particularly acute in the medical image domain, where both the cost of annotation and the inter-observer variability are high. Different human experts contribute estimates of the "actual" segmentation labels in a typical label acquisition process, influenced by their personal biases and competency levels. The performance of automatic segmentation algorithms is limited when these noisy labels are used as the expert consensus label. In this work, we use two coupled CNNs to jointly learn, from purely noisy observations alone, the reliability of individual annotators and the expert consensus label distributions. The separation of the two is achieved by maximally describing the annotator's "unreliable behavior" (we call it "maximally unreliable") while achieving high fidelity with the noisy training data. We first create a toy segmentation dataset using MNIST and investigate the properties of the proposed algorithm. We then use three public medical imaging segmentation datasets to demonstrate our method's efficacy, including both simulated (where necessary) and real-world annotations: 1) ISBI2015 (multiple-sclerosis lesions); 2) BraTS (brain tumors); 3) LIDC-IDRI (lung abnormalities). Finally, we create a real-world multiple sclerosis lesion dataset (QSMSC at UCL: Queen Square Multiple Sclerosis Center at UCL, UK) with manual segmentations from 4 different annotators (3 radiologists with different level skills and 1 expert to generate the expert consensus label). In all datasets, our method consistently outperforms competing methods and relevant baselines, especially when the number of annotations is small and the amount of disagreement is large. The studies also reveal that the system is capable of capturing the complicated spatial characteristics of annotators' mistakes.

Ten years of image analysis and machine learning competitions in dementia.

Machine learning methods exploiting multi-parametric biomarkers, especially based on neuroimaging, have huge potential to improve early diagnosis of dementia and to predict which individuals are at-risk of developing dementia. To benchmark algorithms in the field of machine learning and neuroimaging in dementia and assess their potential for use in clinical practice and clinical trials, seven grand challenges have been organized in the last decade: MIRIAD (2012), Alzheimer's Disease Big Data DREAM (2014), CADDementia (2014), Machine Learning Challenge (2014), MCI Neuroimaging (2017), TADPOLE (2017), and the Predictive Analytics Competition (2019). Based on two challenge evaluation frameworks, we analyzed how these grand challenges are complementing each other regarding research questions, datasets, validation approaches, results and impact. The seven grand challenges addressed questions related to screening, clinical status estimation, prediction and monitoring in (pre-clinical) dementia. There was little overlap in clinical questions, tasks and performance metrics. Whereas this aids providing insight on a broad range of questions, it also limits the validation of results across challenges. The validation process itself was mostly comparable between challenges, using similar methods for ensuring objective comparison, uncertainty estimation and statistical testing. In general, winning algorithms performed rigorous data pre-processing and combined a wide range of input features. Despite high state-of-the-art performances, most of the methods evaluated by the challenges are not clinically used. To increase impact, future challenges could pay more attention to statistical analysis of which factors (i.e., features, models) relate to higher performance, to clinical questions beyond Alzheimer's disease, and to using testing data beyond the Alzheimer's Disease Neuroimaging Initiative. Grand challenges would be an ideal venue for assessing the generalizability of algorithm performance to unseen data of other cohorts. Key for increasing impact in this way are larger testing data sizes, which could be reached by sharing algorithms rather than data to exploit data that cannot be shared. Given the potential and lessons learned in the past ten years, we are excited by the prospects of grand challenges in machine learning and neuroimaging for the next ten years and beyond.

Avoiding Unnecessary Biopsy after Multiparametric Prostate MRI with VERDICT Analysis: The INNOVATE Study.

Background In men suspected of having prostate cancer (PCa), up to 50% of men with positive multiparametric MRI (mpMRI) findings (Prostate Imaging Reporting and Data System [PI-RADS] or Likert score of 3 or higher) have no clinically significant (Gleason score ≤3+3, benign) biopsy findings. Vascular, Extracellular, and Restricted Diffusion for Cytometry in Tumor (VERDICT) MRI analysis could improve the stratification of positive mpMRI findings. Purpose To evaluate VERDICT MRI, mpMRI-derived apparent diffusion coefficient (ADC), and prostate-specific antigen density (PSAD) as determinants of clinically significant PCa (csPCa). Materials and Methods Between April 2016 and December 2019, men suspected of having PCa were prospectively recruited from two centers and underwent VERDICT MRI and mpMRI at one center before undergoing targeted biopsy. Biopsied lesion ADC, lesion-derived fractional intracellular volume (FIC), and PSAD were compared between men with csPCa and those without csPCa, using nonparametric tests subdivided by Likert scores. Area under the receiver operating characteristic curve (AUC) was calculated to test diagnostic performance. Results Among 303 biopsy-naive men, 165 study participants (mean age, 65 years ± 7 [SD]) underwent targeted biopsy; of these, 73 had csPCa. Median lesion FIC was higher in men with csPCa (FIC, 0.53) than in those without csPCa (FIC, 0.18) for Likert 3 (P = .002) and Likert 4 (0.60 vs 0.28, P < .001) lesions. Median lesion ADC was lower for Likert 4 lesions with csPCa (0.86 × 10-3 mm2/sec) compared with lesions without csPCa (1.12 × 10-3 mm2/sec, P = .03), but there was no evidence of a difference for Likert 3 lesions (0.97 × 10-3 mm2/sec vs 1.20 × 10-3 mm2/sec, P = .09). PSAD also showed no difference for Likert 3 (0.17 ng/mL2 vs 0.12 ng/mL2, P = .07) or Likert 4 (0.14 ng/mL2 vs 0.12 ng/mL2, P = .47) lesions. The diagnostic performance of FIC (AUC, 0.96; 95% CI: 0.93, 1.00) was higher (P = .02) than that of ADC (AUC, 0.85; 95% CI: 0.79, 0.91) and PSAD (AUC, 0.74; 95% CI: 0.66, 0.82) for the presence of csPCa in biopsied lesions. Conclusion Lesion fractional intracellular volume enabled better classification of clinically significant prostate cancer than did apparent diffusion coefficient and prostate-specific antigen density. Clinical trial registration no. NCT02689271 © RSNA, 2022 Online supplemental material is available for this article.

Training data distribution significantly impacts the estimation of tissue microstructure with machine learning.

PURPOSE: Supervised machine learning (ML) provides a compelling alternative to traditional model fitting for parameter mapping in quantitative MRI. The aim of this work is to demonstrate and quantify the effect of different training data distributions on the accuracy and precision of parameter estimates when supervised ML is used for fitting. METHODS: We fit a two- and three-compartment biophysical model to diffusion measurements from in-vivo human brain, as well as simulated diffusion data, using both traditional model fitting and supervised ML. For supervised ML, we train several artificial neural networks, as well as random forest regressors, on different distributions of ground truth parameters. We compare the accuracy and precision of parameter estimates obtained from the different estimation approaches using synthetic test data. RESULTS: When the distribution of parameter combinations in the training set matches those observed in healthy human data sets, we observe high precision, but inaccurate estimates for atypical parameter combinations. In contrast, when training data is sampled uniformly from the entire plausible parameter space, estimates tend to be more accurate for atypical parameter combinations but may have lower precision for typical parameter combinations. CONCLUSION: This work highlights that estimation of model parameters using supervised ML depends strongly on the training-set distribution. We show that high precision obtained using ML may mask strong bias, and visual assessment of the parameter maps is not sufficient for evaluating the quality of the estimates.

Sequence of clinical and neurodegeneration events in Parkinson's disease progression.

Dementia is one of the most debilitating aspects of Parkinson's disease. There are no validated biomarkers that can track Parkinson's disease progression, nor accurately identify patients who will develop dementia and when. Understanding the sequence of observable changes in Parkinson's disease in people at elevated risk for developing dementia could provide an integrated biomarker for identifying and managing individuals who will develop Parkinson's dementia. We aimed to estimate the sequence of clinical and neurodegeneration events, and variability in this sequence, using data-driven statistical modelling in two separate Parkinson's cohorts, focusing on patients at elevated risk for dementia due to their age at symptom onset. We updated a novel version of an event-based model that has only recently been extended to cope naturally with clinical data, enabling its application in Parkinson's disease for the first time. The observational cohorts included healthy control subjects and patients with Parkinson's disease, of whom those diagnosed at age 65 or older were classified as having high risk of dementia. The model estimates that Parkinson's progression in patients at elevated risk for dementia starts with classic prodromal features of Parkinson's disease (olfaction, sleep), followed by early deficits in visual cognition and increased brain iron content, followed later by a less certain ordering of neurodegeneration in the substantia nigra and cortex, neuropsychological cognitive deficits, retinal thinning in dopamine layers, and further deficits in visual cognition. Importantly, we also characterize variation in the sequence. We found consistent, cross-validated results within cohorts, and agreement between cohorts on the subset of features available in both cohorts. Our sequencing results add powerful support to the increasing body of evidence suggesting that visual processing specifically is affected early in patients with Parkinson's disease at elevated risk of dementia. This opens a route to earlier and more precise detection, as well as a more detailed understanding of the pathological mechanisms underpinning Parkinson's dementia.

Applying causal models to explore the mechanism of action of simvastatin in progressive multiple sclerosis.

Understanding the mode of action of drugs is a challenge with conventional methods in clinical trials. Here, we aimed to explore whether simvastatin effects on brain atrophy and disability in secondary progressive multiple sclerosis (SPMS) are mediated by reducing cholesterol or are independent of cholesterol. We applied structural equation models to the MS-STAT trial in which 140 patients with SPMS were randomized to receive placebo or simvastatin. At baseline, after 1 and 2 years, patients underwent brain magnetic resonance imaging; their cognitive and physical disability were assessed on the block design test and Expanded Disability Status Scale (EDSS), and serum total cholesterol levels were measured. We calculated the percentage brain volume change (brain atrophy). We compared two models to select the most likely one: a cholesterol-dependent model with a cholesterol-independent model. The cholesterol-independent model was the most likely option. When we deconstructed the total treatment effect into indirect effects, which were mediated by brain atrophy, and direct effects, simvastatin had a direct effect (independent of serum cholesterol) on both the EDSS, which explained 69% of the overall treatment effect on EDSS, and brain atrophy, which, in turn, was responsible for 31% of the total treatment effect on EDSS [ß = -0.037; 95% credible interval (CI) = -0.075, -0.010]. This suggests that simvastatin's beneficial effects in MS are independent of its effect on lowering peripheral cholesterol levels, implicating a role for upstream intermediate metabolites of the cholesterol synthesis pathway. Importantly, it demonstrates that computational models can elucidate the causal architecture underlying treatment effects in clinical trials of progressive MS.

On the potential for mapping apparent neural soma density via a clinically viable diffusion MRI protocol.

Diffusion MRI is a valuable tool for probing tissue microstructure in the brain noninvasively. Today, model-based techniques are widely available and used for white matter characterisation where their development is relatively mature. Conversely, tissue modelling in grey matter is more challenging, and no generally accepted models exist. With advances in measurement technology and modelling efforts, a clinically viable technique that reveals salient features of grey matter microstructure, such as the density of quasi-spherical cell bodies and quasi-cylindrical cell projections, is an exciting prospect. As a step towards capturing the microscopic architecture of grey matter in clinically feasible settings, this work uses a biophysical model that is designed to disentangle the diffusion signatures of spherical and cylindrical structures in the presence of orientation heterogeneity, and takes advantage of B-tensor encoding measurements, which provide additional sensitivity compared to standard single diffusion encoding sequences. For the fast and robust estimation of microstructural parameters, we leverage recent advances in machine learning and replace conventional fitting techniques with an artificial neural network that fits complex biophysical models within seconds. Our results demonstrate apparent markers of spherical and cylindrical geometries in healthy human subjects, and in particular an increased volume fraction of spherical compartments in grey matter compared to white matter. We evaluate the extent to which spherical and cylindrical geometries may be interpreted as correlates of neural soma and neural projections, respectively, and quantify parameter estimation errors in the presence of various departures from the modelling assumptions. While further work is necessary to translate the ideas presented in this work to the clinic, we suggest that biomarkers focussing on quasi-spherical cellular geometries may be valuable for the enhanced assessment of neurodevelopmental disorders and neurodegenerative diseases.

Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI.

Deep learning (DL) has shown great potential in medical image enhancement problems, such as super-resolution or image synthesis. However, to date, most existing approaches are based on deterministic models, neglecting the presence of different sources of uncertainty in such problems. Here we introduce methods to characterise different components of uncertainty, and demonstrate the ideas using diffusion MRI super-resolution. Specifically, we propose to account for intrinsic uncertainty through a heteroscedastic noise model and for parameter uncertainty through approximate Bayesian inference, and integrate the two to quantify predictive uncertainty over the output image. Moreover, we introduce a method to propagate the predictive uncertainty on a multi-channelled image to derived scalar parameters, and separately quantify the effects of intrinsic and parameter uncertainty therein. The methods are evaluated for super-resolution of two different signal representations of diffusion MR images-Diffusion Tensor images and Mean Apparent Propagator MRI-and their derived quantities such as mean diffusivity and fractional anisotropy, on multiple datasets of both healthy and pathological human brains. Results highlight three key potential benefits of modelling uncertainty for improving the safety of DL-based image enhancement systems. Firstly, modelling uncertainty improves the predictive performance even when test data departs from training data ("out-of-distribution" datasets). Secondly, the predictive uncertainty highly correlates with reconstruction errors, and is therefore capable of detecting predictive "failures". Results on both healthy subjects and patients with brain glioma or multiple sclerosis demonstrate that such an uncertainty measure enables subject-specific and voxel-wise risk assessment of the super-resolved images that can be accounted for in subsequent analysis. Thirdly, we show that the method for decomposing predictive uncertainty into its independent sources provides high-level "explanations" for the model performance by separately quantifying how much uncertainty arises from the inherent difficulty of the task or the limited training examples. The introduced concepts of uncertainty modelling extend naturally to many other imaging modalities and data enhancement applications.

Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast.

Most existing algorithms for automatic 3D morphometry of human brain MRI scans are designed for data with near-isotropic voxels at approximately 1 mm resolution, and frequently have contrast constraints as well-typically requiring T1-weighted images (e.g., MP-RAGE scans). This limitation prevents the analysis of millions of MRI scans acquired with large inter-slice spacing in clinical settings every year. In turn, the inability to quantitatively analyze these scans hinders the adoption of quantitative neuro imaging in healthcare, and also precludes research studies that could attain huge sample sizes and hence greatly improve our understanding of the human brain. Recent advances in convolutional neural networks (CNNs) are producing outstanding results in super-resolution and contrast synthesis of MRI. However, these approaches are very sensitive to the specific combination of contrast, resolution and orientation of the input images, and thus do not generalize to diverse clinical acquisition protocols - even within sites. In this article, we present SynthSR, a method to train a CNN that receives one or more scans with spaced slices, acquired with different contrast, resolution and orientation, and produces an isotropic scan of canonical contrast (typically a 1 mm MP-RAGE). The presented method does not require any preprocessing, beyond rigid coregistration of the input scans. Crucially, SynthSR trains on synthetic input images generated from 3D segmentations, and can thus be used to train CNNs for any combination of contrasts, resolutions and orientations without high-resolution real images of the input contrasts. We test the images generated with SynthSR in an array of common downstream analyses, and show that they can be reliably used for subcortical segmentation and volumetry, image registration (e.g., for tensor-based morphometry), and, if some image quality requirements are met, even cortical thickness morphometry. The source code is publicly available at https://github.com/BBillot/SynthSR.