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.

Tertiary lymphoid structures (TLS) identification and density assessment on H&E-stained digital slides of lung cancer.

Tertiary lymphoid structures (TLS) are ectopic aggregates of lymphoid cells in inflamed, infected, or tumoral tissues that are easily recognized on an H&E histology slide as discrete entities, distinct from lymphocytes. TLS are associated with improved cancer prognosis but there is no standardised method available to quantify their presence. Previous studies have used immunohistochemistry to determine the presence of specific cells as a marker of the TLS. This has now been proven to be an underestimate of the true number of TLS. Thus, we propose a methodology for the automated identification and quantification of TLS, based on H&E slides. We subsequently determined the mathematical criteria defining a TLS. TLS regions were identified through a deep convolutional neural network and segmentation of lymphocytes was performed through an ellipsoidal model. This methodology had a 92.87% specificity at 95% sensitivity, 88.79% specificity at 98% sensitivity and 84.32% specificity at 99% sensitivity level based on 144 TLS annotated H&E slides implying that the automated approach was able to reproduce the histopathologists' assessment with great accuracy. We showed that the minimum number of lymphocytes within TLS is 45 and the minimum TLS area is 6,245µm2. Furthermore, we have shown that the density of the lymphocytes is more than 3 times those outside of the TLS. The mean density and standard deviation of lymphocytes within a TLS area are 0.0128/µm2 and 0.0026/µm2 respectively compared to 0.004/µm2 and 0.001/µm2 in non-TLS regions. The proposed methodology shows great potential for automated identification and quantification of the TLS density on digital H&E slides.

Pleuroparenchymal fibroelastosis in idiopathic pulmonary fibrosis: Survival analysis using visual and computer-based computed tomography assessment.

BACKGROUND: Idiopathic pulmonary fibrosis (IPF) and pleuroparenchymal fibroelastosis (PPFE) are known to have poor outcomes but detailed examinations of prognostic significance of an association between these morphologic processes are lacking. METHODS: Retrospective observational study of independent derivation and validation cohorts of IPF populations. Upper-lobe PPFE extent was scored visually (vPPFE) as categories of absent, moderate, marked. Computerised upper-zone PPFE extent (cPPFE) was examined continuously and using a threshold of 2·5% pleural surface area. vPPFE and cPPFE were evaluated against 1-year FVC decline (estimated using mixed-effects models) and mortality. Multivariable models were adjusted for age, gender, smoking history, antifibrotic treatment and diffusion capacity for carbon monoxide. FINDINGS: PPFE prevalence was 49% (derivation cohort, n = 142) and 72% (validation cohort, n = 145). vPPFE marginally contributed 3-14% to variance in interstitial lung disease (ILD) severity across both cohorts.In multivariable models, marked vPPFE was independently associated with 1-year FVC decline (derivation: regression coefficient 18·3, 95 CI 8·47-28·2%; validation: 7·51, 1·85-13·2%) and mortality (derivation: hazard ratio [HR] 7·70, 95% CI 3·50-16·9; validation: HR 3·01, 1·33-6·81). Similarly, continuous and dichotomised cPPFE were associated with 1-year FVC decline and mortality (cPPFE ≥ 2·5% derivation: HR 5·26, 3·00-9·22; validation: HR 2·06, 1·28-3·31). Individuals with cPPFE ≥ 2·5% or marked vPPFE had the lowest median survival, the cPPFE threshold demonstrated greater discrimination of poor outcomes at two and three years than marked vPPFE. INTERPRETATION: PPFE quantification supports distinction of IPF patients with a worse outcome independent of established ILD severity measures. This has the potential to improve prognostic management and elucidate separate pathways of disease progression. FUNDING: This research was funded in whole or in part by the Wellcome Trust [209,553/Z/17/Z] and the NIHR UCLH Biomedical Research Centre, UK.

Machine learning based white matter models with permeability: An experimental study in cuprizone treated in-vivo mouse model of axonal demyelination.

The intra-axonal water exchange time (τi), a parameter associated with axonal permeability, could be an important biomarker for understanding and treating demyelinating pathologies such as Multiple Sclerosis. Diffusion-Weighted MRI (DW-MRI) is sensitive to changes in permeability; however, the parameter has so far remained elusive due to the lack of general biophysical models that incorporate it. Machine learning based computational models can potentially be used to estimate such parameters. Recently, for the first time, a theoretical framework using a random forest (RF) regressor suggests that this is a promising new approach for permeability estimation. In this study, we adopt such an approach and for the first time experimentally investigate it for demyelinating pathologies through direct comparison with histology. We construct a computational model using Monte Carlo simulations and an RF regressor in order to learn a mapping between features derived from DW-MRI signals and ground truth microstructure parameters. We test our model in simulations, and find strong correlations between the predicted and ground truth parameters (intra-axonal volume fraction f: R2 =0.99, τi: R2 =0.84, intrinsic diffusivity d: R2 =0.99). We then apply the model in-vivo, on a controlled cuprizone (CPZ) mouse model of demyelination, comparing the results from two cohorts of mice, CPZ (N=8) and healthy age-matched wild-type (WT, N=8). We find that the RF model estimates sensible microstructure parameters for both groups, matching values found in literature. Furthermore, we perform histology for both groups using electron microscopy (EM), measuring the thickness of the myelin sheath as a surrogate for exchange time. Histology results show that our RF model estimates are very strongly correlated with the EM measurements (ρ = 0.98 for f, ρ = 0.82 for τi). Finally, we find a statistically significant decrease in τi in all three regions of the corpus callosum (splenium/genu/body) of the CPZ cohort (<τi>=310ms/330ms/350ms) compared to the WT group (<τi>=370ms/370ms/380ms). This is in line with our expectations that τi is lower in regions where the myelin sheath is damaged, as axonal membranes become more permeable. Overall, these results demonstrate, for the first time experimentally and in vivo, that a computational model learned from simulations can reliably estimate microstructure parameters, including the axonal permeability .

Improving the characterization of meningioma microstructure in proton therapy from conventional apparent diffusion coefficient measurements using Monte Carlo simulations of diffusion MRI.

PURPOSE: Proton therapy could benefit from noninvasively gaining tumor microstructure information, at both planning and monitoring stages. The anatomical location of brain tumors, such as meningiomas, often hinders the recovery of such information from histopathology, and conventional noninvasive imaging biomarkers, like the apparent diffusion coefficient (ADC) from diffusion-weighted MRI (DW-MRI), are nonspecific. The aim of this study was to retrieve discriminative microstructural markers from conventional ADC for meningiomas treated with proton therapy. These markers were employed for tumor grading and tumor response assessment. METHODS: DW-MRIs from patients affected by meningioma and enrolled in proton therapy were collected before (n = 35) and 3 months after (n = 25) treatment. For the latter group, the risk of an adverse outcome was inferred by their clinical history. Using Monte Carlo methods, DW-MRI signals were simulated from packings of synthetic cells built with well-defined geometrical and diffusion properties. Patients' ADC was modeled as a weighted sum of selected simulated signals. The weights that best described a patient's ADC were determined through an optimization procedure and used to estimate a set of markers of tumor microstructure: diffusion coefficient (D), volume fraction (vf), and radius (R). Apparent cellularity (ρapp ) was estimated from vf and R for an easier clinical interpretability. Differences between meningothelial and atypical subtypes, and low- and high-grade meningiomas were assessed with nonparametric statistical tests, whereas sensitivity and specificity with ROC analyses. Similar analyses were performed for patients showing low or high risk of an adverse outcome to preliminary evaluate response to treatment. RESULTS: Significant (P < 0.05) differences in median ADC, D, vf, R, and ρapp values were found when comparing meningiomas' subtypes and grades. ROC analyses showed that estimated microstructural parameters reached higher specificity than ADC for subtyping (0.93 for D and vf vs 0.80 for ADC) and grading (0.75 for R vs 0.67 for ADC). High- and low-risk patients showed significant differences in ADC and microstructural parameters. The skewness of ρapp was the parameter with highest AUC (0.90) and sensitivity (0.75). CONCLUSIONS: Matching measured with simulated ADC yielded a set of potential imaging markers for meningiomas grading and response monitoring in proton therapy, showing higher specificity than conventional ADC. These markers can provide discriminative information about spatial patterns of tumor microstructure implying important advantages for patient-specific proton therapy workflows.

Augmenting Dementia Cognitive Assessment With Instruction-Less Eye-Tracking Tests.

Eye-tracking technology is an innovative tool that holds promise for enhancing dementia screening. In this work, we introduce a novel way of extracting salient features directly from the raw eye-tracking data of a mixed sample of dementia patients during a novel instruction-less cognitive test. Our approach is based on self-supervised representation learning where, by training initially a deep neural network to solve a pretext task using well-defined available labels (e.g. recognising distinct cognitive activities in healthy individuals), the network encodes high-level semantic information which is useful for solving other problems of interest (e.g. dementia classification). Inspired by previous work in explainable AI, we use the Layer-wise Relevance Propagation (LRP) technique to describe our network's decisions in differentiating between the distinct cognitive activities. The extent to which eye-tracking features of dementia patients deviate from healthy behaviour is then explored, followed by a comparison between self-supervised and handcrafted representations on discriminating between participants with and without dementia. Our findings not only reveal novel self-supervised learning features that are more sensitive than handcrafted features in detecting performance differences between participants with and without dementia across a variety of tasks, but also validate that instruction-less eye-tracking tests can detect oculomotor biomarkers of dementia-related cognitive dysfunction. This work highlights the contribution of self-supervised representation learning techniques in biomedical applications where the small number of patients, the non-homogenous presentations of the disease and the complexity of the setting can be a challenge using state-of-the-art feature extraction methods.

Relevance of time-dependence for clinically viable diffusion imaging of the spinal cord.

PURPOSE: Time-dependence is a key feature of the diffusion-weighted (DW) signal, knowledge of which informs biophysical modelling. Here, we study time-dependence in the human spinal cord, as its axonal structure is specific and different from the brain. METHODS: We run Monte Carlo simulations using a synthetic model of spinal cord white matter (WM) (large axons), and of brain WM (smaller axons). Furthermore, we study clinically feasible multi-shell DW scans of the cervical spinal cord (b = 0; b = 711 s mm-2 ; b = 2855 s mm-2 ), obtained using three diffusion times (Δ of 29, 52 and 76 ms) from three volunteers. RESULTS: Both intra-/extra-axonal perpendicular diffusivities and kurtosis excess show time-dependence in our synthetic spinal cord model. This time-dependence is reflected mostly in the intra-axonal perpendicular DW signal, which also exhibits strong decay, unlike our brain model. Time-dependence of the total DW signal appears detectable in the presence of noise in our synthetic spinal cord model, but not in the brain. In WM in vivo, we observe time-dependent macroscopic and microscopic diffusivities and diffusion kurtosis, NODDI and two-compartment SMT metrics. Accounting for large axon calibers improves fitting of multi-compartment models to a minor extent. CONCLUSIONS: Time-dependence of clinically viable DW MRI metrics can be detected in vivo in spinal cord WM, thus providing new opportunities for the non-invasive estimation of microstructural properties. The time-dependence of the perpendicular DW signal may feature strong intra-axonal contributions due to large spinal axon caliber. Hence, a popular model known as "stick" (zero-radius cylinder) may be sub-optimal to describe signals from the largest spinal axons.

Machine learning based compartment models with permeability for white matter microstructure imaging.

Some microstructure parameters, such as permeability, remain elusive because mathematical models that express their relationship to the MR signal accurately are intractable. Here, we propose to use computational models learned from simulations to estimate these parameters. We demonstrate the approach in an example which estimates water residence time in brain white matter. The residence time τi of water inside axons is a potentially important biomarker for white matter pathologies of the human central nervous system, as myelin damage is hypothesised to affect axonal permeability, and thus τi. We construct a computational model using Monte Carlo simulations and machine learning (specifically here a random forest regressor) in order to learn a mapping between features derived from diffusion weighted MR signals and ground truth microstructure parameters, including τi. We test our numerical model using simulated and in vivo human brain data. Simulation results show that estimated parameters have strong correlations with the ground truth parameters (R2={0.88,0.95,0.82,0.99}) for volume fraction, residence time, axon radius and diffusivity respectively), and provide a marked improvement over the most widely used Kärger model (R2={0.75,0.60,0.11,0.99}). The trained model also estimates sensible microstructure parameters from in vivo human brain data acquired from healthy controls, matching values found in literature, and provides better reproducibility than the Kärger model on both the voxel and ROI level. Finally, we acquire data from two Multiple Sclerosis (MS) patients and compare to the values in healthy subjects. We find that in the splenium of corpus callosum (CC-S) the estimate of the residence time is 0.57±0.05s for the healthy subjects, while in the MS patient with a lesion in CC-S it is 0.33±0.12s in the normal appearing white matter (NAWM) and 0.19±0.11s in the lesion. In the corticospinal tracts (CST) the estimate of the residence time is 0.52±0.09s for the healthy subjects, while in the MS patient with a lesion in CST it is 0.56±0.05s in the NAWM and 0.13±0.09s in the lesion. These results agree with our expectations that the residence time in lesions would be lower than in NAWM because the loss of myelin should increase permeability. Overall, we find parameter estimates in the two MS patients consistent with expectations from the pathology of MS lesions demonstrating the clinical potential of this new technique.

Microstructural parameter estimation in vivo using diffusion MRI and structured prior information.

PURPOSE: Diffusion MRI has recently been used with detailed models to probe tissue microstructure. Much of this work has been performed ex vivo with powerful scanner hardware, to gain sensitivity to parameters such as axon radius. By contrast, performing microstructure imaging on clinical scanners is extremely challenging. METHODS: We use an optimized dual spin-echo diffusion protocol, and a Bayesian fitting approach, to obtain reproducible contrast (histogram overlap of up to 92%) in estimated maps of axon radius index in healthy adults at a modest, widely-available gradient strength (35 mT m(-1)). A key innovation is the use of influential priors. RESULTS: We demonstrate that our priors can improve precision in axon radius estimates--a 7-fold reduction in voxelwise coefficient of variation in vivo--without significant bias. Our results may reflect true axon radius differences between white matter regions, but this interpretation should be treated with caution due to the complexity of the tissue relative to our model. CONCLUSIONS: Some sensitivity to relatively large axons (3-15 µm) may be available at clinical field and gradient strengths. Future applications at higher gradient strength will benefit from the favorable eddy current properties of the dual spin-echo sequence, and greater precision available with suitable priors. Magnetic Resonance in Medicine published by Wiley Periodicals, Inc. on behalf of International Society for Magnetic Resonance.

Multimodality characterization of microstructure by the combination of diffusion NMR and time-domain diffuse optical data.

Combining datasets with a model of the underlying physics prior to mapping of tissue provides a novel approach improving the estimation of parameters. We demonstrate this approach by merging near infrared diffuse optical signal data with diffusion NMR data to inform a model describing the microstructure of a sample. The study is conducted on a homogeneous emulsion of oil in a dispersion medium of water and proteins. The use of a protein based background, rich in collagen, introduces a similarity to real tissues compared to other models such as intralipids. The sample is investigated with the two modalities separately. Then, the two datasets are used to inform a combined model, and to estimate the size of the microstructural elements and the volume fraction. The combined model fits the microstructural properties by minimizing the difference between experimental and modelled data. The experimental results are validated with confocal laser scanning microscopy. The final results demonstrate that the combined model provides improved estimates of microstructural parameters compared to either individual model alone.

Contrast and stability of the axon diameter index from microstructure imaging with diffusion MRI.

The ActiveAx technique fits the minimal model of white matter diffusion to diffusion MRI data acquired using optimized protocols that provide orientationally invariant indices of axon diameter and density. We investigated how limitations of the available maximal gradient strength (Gmax) on a scanner influence the sensitivity to a range of axon diameters. Multishell high-angular-diffusion-imaging (HARDI) protocols for Gmax of 60, 140, 200, and 300 mT/m were optimized for the pulsed-gradient-spin-echo (PGSE) sequence. Data were acquired on a fixed monkey brain and Monte-Carlo simulations supported the results. Increasing Gmax reduces within-voxel variation of the axon diameter index and improves contrast beyond what is achievable with higher signal-to-noise ratio. Simulations reveal an upper bound on the axon diameter (∼10 µm) that pulsed-gradient-spin-echo measurements are sensitive to, due to a trade-off between short T2 and the long diffusion time needed to probe larger axon diameters. A lower bound (∼2.5 µm) slightly dependent on Gmax was evident, below which axon diameters are identifiable as small, but impossible to differentiate. These results emphasize the key-role of Gmax for enhancing contrast between axon diameter distributions and are, therefore, relevant in general for microstructure imaging methods and highlight the need for increased Gmax on future commercial systems.

A new approach to structural integrity assessment based on axial and radial diffusivities.

This study describes a method for performing diffusivity measures along and across a specific direction, derived from white matter in healthy controls. The diffusion tensor (DT) assigns a principal eigenvector (v1) and eigenvalue (axial diffusivity, d(ax)) to each voxel. The average of the second and third eigenvalues is the radial diffusivity, d(rad) v1 may be affected by pathology, therefore when comparing d(ax) and d(rad) in patients one has to consider the direction of the measurement and underlying anatomy. Here we created a representative super-DT dataset, DT(ref), whose eigenvector, v(1,ref), was considered the most likely direction of diffusivity per voxel. We defined the projected axial diffusivity, d(p-ax), as the projection of individual DTs along v(1,ref) and the projected radial diffusivity, d(p-rad), as the average of the projections along the second and third eigenvectors of DT(ref). The projected diffusivities are promising new parameters for studying white matter pathology.

Axon diameter mapping in crossing fibers with diffusion MRI.

This paper proposes a technique for a previously unaddressed problem, namely, mapping axon diameter in crossing fiber regions, using diffusion MRI. Direct measurement of tissue microstructure of this kind using diffusion MRI offers a new class of biomarkers that give more specific information about tissue than measures derived from diffusion tensor imaging. Most existing techniques for axon diameter mapping assume a single axon orientation in the tissue model, which limits their application to only the most coherently oriented brain white matter, such as the corpus callosum, where the single orientation assumption is a reasonable one. However, fiber crossings and other complex configurations are widespread in the brain. In such areas, the existing techniques will fail to provide useful axon diameter indices for any of the individual fiber populations. We propose a novel crossing fiber tissue model to enable axon diameter mapping in voxels with crossing fibers. We show in simulation that the technique can provide robust axon diameter estimates in a two-fiber crossing with the crossing angle as small as 45 degrees. Using ex vivo imaging data, we further demonstrate the feasibility of the technique by establishing reasonable axon diameter indices in the crossing region at the interface of the cingulum and the corpus callosum.

The matrix formalism for generalised gradients with time-varying orientation in diffusion NMR.

The matrix formalism is a general framework for evaluating the diffusion NMR signal from restricted spins under generalised gradient waveforms. The original publications demonstrate the method for waveforms that vary only in magnitude and have fixed orientation. In this work, we extend the method to allow for variations in the direction of the gradient. This extension is necessary, for example to incorporate the effects of crusher gradients or imaging gradients in diffusion MRI, to characterise signal anisotropy in double pulsed field gradient (dPFG) experiments, or to optimise the gradient waveform for microstructure sensitivity. In particular, we show for primitive geometries (planes, cylinders and spheres), how to express the matrix operators at each time point of the gradient waveform as a linear combination of one or two fundamental matrices. Thus we obtain an efficient implementation with both the storage and CPU demands similar to the fixed-orientation case. Comparison with Monte Carlo simulations validates the implementation on three different sequences: dPFG, helical waveforms and the stimulated echo (STEAM) sequence.

Convergence and parameter choice for Monte-Carlo simulations of diffusion MRI.

This paper describes a general and flexible Monte- Carlo simulation framework for diffusing spins that generates realistic synthetic data for diffusion magnetic resonance imaging. Similar systems in the literature consider only simple substrates and their authors do not consider convergence and parameter optimization. We show how to run Monte-Carlo simulations within complex irregular substrates. We compare the results of the Monte-Carlo simulation to an analytical model of restricted diffusion to assess precision and accuracy of the generated results. We obtain an optimal combination of spins and updates for a given run time by trading off number of updates in favor of number of spins such that precision and accuracy of sythesized data are both optimized. Further experiments demonstrate the system using a tissue environment that current analytic models cannot capture. This tissue model incorporates swelling, abutting, and deformation. Swelling-induced restriction in the extracellular space due to the effects of abutting cylinders leads to large departures from the predictions of the analytical model, which does not capture these effects. This swelling-induced restriction may be an important mechanism in explaining the changes in apparent diffusion constant observed in the aftermath of acute ischemic stroke.

Optimal acquisition schemes for in vivo quantitative magnetization transfer MRI.

This paper uses the theory of Cramer-Rao lower bounds (CRLB) to obtain optimal acquisition schemes for in vivo quantitative magnetization transfer (MT) imaging, although the method is generally applicable to any multiparametric MRI technique. Quantitative MT fits a two-pool model to data collected at different sampling points or settings of amplitude and offset frequency in the MT saturation pulses. Here we use simple objective functions based on the CRLB to optimize sampling strategies for multiple parameters simultaneously, and use simulated annealing to minimize these objective functions with respect to the sampling configuration. Experiments compare optimal schemes derived for quantitative MT in the human white matter (WM) at 1.5T with previously published schemes using both synthetic and human-brain data. Results show large reductions in error of the fitted parameters with the new schemes, which greatly increases the clinical potential of in vivo quantitative MT. Since the sampling-scheme optimization requires specific settings of the MT parameters, we also show that the optimum schemes are robust to these settings within the range of MT parameters observed in the brain.

Optimal imaging parameters for fiber-orientation estimation in diffusion MRI.

This study uses Monte Carlo simulations to investigate the optimal value of the diffusion weighting factor b for estimating white-matter fiber orientations using diffusion MRI with a standard spherical sampling scheme. We devise an algorithm for determining the optimal echo time, pulse width, and pulse separation in the pulsed-gradient spin-echo sequence for a specific value of b. The Monte Carlo simulations provide an estimate of the optimal value of b for recovering one and two fiber orientations. We show that the optimum is largely independent of the noise level in the measurements and the number of gradient directions and that the optimum depends only weakly on the diffusion anisotropy, the maximum gradient strength, and the spin-spin relaxation time. The optimum depends strongly on the mean diffusivity. In brain tissue, the optima we estimate are in the ranges [0.7, 1.0] x 10(9) s m(-2) and [2.2, 2.8] x 10(9) s m(-2) for the one- and two-fiber cases, respectively. The best b for estimating the fractional anisotropy is slightly higher than for estimating fiber directions in the one-fiber case and slightly lower in the two-fiber case. To estimate Tr(D) in the one-fiber case, the optimal setting is higher still. Simulations suggest that a ratio of high to low b measurements of 5 to 1 is a good compromise for measuring fiber directions and size and shape indices.

Probabilistic Monte Carlo based mapping of cerebral connections utilising whole-brain crossing fibre information.

A methodology is presented for estimation of a probability density function of cerebral fibre orientations when one or two fibres are present in a voxel. All data are acquired on a clinical MR scanner, using widely available acquisition techniques. The method models measurements of water diffusion in a single fibre by a Gaussian density function and in multiple fibres by a mixture of Gaussian densities. The effects of noise on complex MR diffusion weighted data are explicitly simluated and parameterised. This information is used for standard and Monte Carlo streamline methods. Deterministic and probabilistic maps of anatomical voxel scale connectivity between brain regions are generated.