Accuracy of Automated Computer-Aided Diagnosis for Stroke Imaging: A Critical Evaluation of Current Evidence.

There is increasing interest in computer applications, using artificial intelligence methodologies, to perform health care tasks previously performed by humans, particularly in medical imaging for diagnosis. In stroke, there are now commercial artificial intelligence software for use with computed tomography or MR imaging to identify acute ischemic brain tissue pathology, arterial obstruction on computed tomography angiography or as hyperattenuated arteries on computed tomography, brain hemorrhage, or size of perfusion defects. A rapid, accurate diagnosis may aid treatment decisions for individual patients and could improve outcome if it leads to effective and safe treatment; or conversely, to disaster if a delayed or incorrect diagnosis results in inappropriate treatment. Despite this potential clinical impact, diagnostic tools including artificial intelligence methods are not subjected to the same clinical evaluation standards as are mandatory for drugs. Here, we provide an evidence-based review of the pros and cons of commercially available automated methods for medical imaging diagnosis, including those based on artificial intelligence, to diagnose acute brain pathology on computed tomography or magnetic resonance imaging in patients with stroke.

Reduced structural connectivity within a prefrontal-motor-subcortical network in amyotrophic lateral sclerosis.

BACKGROUND: To investigate white matter structural connectivity changes associated with amyotrophic lateral sclerosis (ALS) using network analysis and compare the results with those obtained using standard voxel-based methods, specifically Tract-based Spatial Statistics (TBSS). METHODS: MRI data were acquired from 30 patients with ALS and 30 age-matched healthy controls. For each subject, 85 grey matter regions (network nodes) were identified from high resolution structural MRI, and network connections formed from the white matter tracts generated by diffusion MRI and probabilistic tractography. Whole-brain networks were constructed using strong constraints on anatomical plausibility and a weighting reflecting tract-averaged fractional anisotropy (FA). RESULTS: Analysis using Network-based Statistics (NBS), without a priori selected regions, identified an impaired motor-frontal-subcortical subnetwork (10 nodes and 12 bidirectional connections), consistent with upper motor neuron pathology, in the ALS group compared with the controls (P = 0.020). Reduced FA in three of the impaired network connections, which involved fibers of the corticospinal tract, correlated with rate of disease progression (P ≤ 0.024). A novel network-tract comparison revealed that the connections involved in the affected network had a strong correspondence (mean overlap of 86.2%) with white matter tracts identified as having reduced FA compared with the control group using TBSS. CONCLUSION: These findings suggest that white matter degeneration in ALS is strongly linked to the motor cortex, and that impaired structural networks identified using NBS have a strong correspondence to affected white matter tracts identified using more conventional voxel-based methods.

Test-retest reliability of structural brain networks from diffusion MRI.

Structural brain networks constructed from diffusion MRI (dMRI) and tractography have been demonstrated in healthy volunteers and more recently in various disorders affecting brain connectivity. However, few studies have addressed the reproducibility of the resulting networks. We measured the test-retest properties of such networks by varying several factors affecting network construction using ten healthy volunteers who underwent a dMRI protocol at 1.5T on two separate occasions. Each T1-weighted brain was parcellated into 84 regions-of-interest and network connections were identified using dMRI and two alternative tractography algorithms, two alternative seeding strategies, a white matter waypoint constraint and three alternative network weightings. In each case, four common graph-theoretic measures were obtained. Network properties were assessed both node-wise and per network in terms of the intraclass correlation coefficient (ICC) and by comparing within- and between-subject differences. Our findings suggest that test-retest performance was improved when: 1) seeding from white matter, rather than grey; and 2) using probabilistic tractography with a two-fibre model and sufficient streamlines, rather than deterministic tensor tractography. In terms of network weighting, a measure of streamline density produced better test-retest performance than tract-averaged diffusion anisotropy, although it remains unclear which is a more accurate representation of the underlying connectivity. For the best performing configuration, the global within-subject differences were between 3.2% and 11.9% with ICCs between 0.62 and 0.76. The mean nodal within-subject differences were between 5.2% and 24.2% with mean ICCs between 0.46 and 0.62. For 83.3% (70/84) of nodes, the within-subject differences were smaller than between-subject differences. Overall, these findings suggest that whilst current techniques produce networks capable of characterising the genuine between-subject differences in connectivity, future work must be undertaken to improve network reliability.

Charles Bonnet syndrome: evidence for a generative model in the cortex?

Several theories propose that the cortex implements an internal model to explain, predict, and learn about sensory data, but the nature of this model is unclear. One condition that could be highly informative here is Charles Bonnet syndrome (CBS), where loss of vision leads to complex, vivid visual hallucinations of objects, people, and whole scenes. CBS could be taken as indication that there is a generative model in the brain, specifically one that can synthesise rich, consistent visual representations even in the absence of actual visual input. The processes that lead to CBS are poorly understood. Here, we argue that a model recently introduced in machine learning, the deep Boltzmann machine (DBM), could capture the relevant aspects of (hypothetical) generative processing in the cortex. The DBM carries both the semantics of a probabilistic generative model and of a neural network. The latter allows us to model a concrete neural mechanism that could underlie CBS, namely, homeostatic regulation of neuronal activity. We show that homeostatic plasticity could serve to make the learnt internal model robust against e.g. degradation of sensory input, but overcompensate in the case of CBS, leading to hallucinations. We demonstrate how a wide range of features of CBS can be explained in the model and suggest a potential role for the neuromodulator acetylcholine. This work constitutes the first concrete computational model of CBS and the first application of the DBM as a model in computational neuroscience. Our results lend further credence to the hypothesis of a generative model in the brain.

Single subject fMRI test-retest reliability metrics and confounding factors.

While the fMRI test-retest reliability has been mainly investigated from the point of view of group level studies, here we present analyses and results for single-subject test-retest reliability. One important aspect of group level reliability is that not only does it depend on between-session variance (test-retest), but also on between-subject variance. This has partly led to a debate regarding which reliability metric to use and how different sources of noise contribute to between-session variance. Focusing on single subject reliability allows considering between-session only. In this study, we measured test-retest reliability in four behavioural tasks (motor mapping, covert verb generation, overt word repetition, and a landmark identification task) to ensure generalisation of the results and at three levels of data processing (time-series correlation, t value variance, and overlap of thresholded maps) to understand how each step influences the other and how confounding factors influence reliability at each of these steps. The contributions of confounding factors (scanner noise, subject motion, and coregistration) were investigated using multiple regression and relative importance analyses at each step. Finally, to achieve a fuller picture of what constitutes a reliable task, we introduced a bootstrap technique of within- vs. between-subject variance. Our results show that (i) scanner noise and coregistration errors have little contribution to between-session variance (ii) subject motion (especially correlated with the stimuli) can have detrimental effects on reliability (iii) different tasks lead to different reliability results. This suggests that between-session variance in fMRI is mostly caused by the variability of underlying cognitive processes and motion correlated with the stimuli rather than technical limitations of data processing.

Can deep learning on retinal images augment known risk factors for cardiovascular disease prediction in diabetes? A prospective cohort study from the national screening programme in Scotland.

AIMS: This study's objective was to evaluate whether deep learning (DL) on retinal photographs from a diabetic retinopathy screening programme improve prediction of incident cardiovascular disease (CVD). METHODS: DL models were trained to jointly predict future CVD risk and CVD risk factors and used to output a DL score. Poisson regression models including clinical risk factors with and without a DL score were fitted to study cohorts with 2,072 and 38,730 incident CVD events in type 1 (T1DM) and type 2 diabetes (T2DM) respectively. RESULTS: DL scores were independently associated with incident CVD with adjusted standardised incidence rate ratios of 1.14 (P = 3 × 10-04 95 % CI (1.06, 1.23)) and 1.16 (P = 4 × 10-33 95 % CI (1.13, 1.18)) in T1DM and T2DM cohorts respectively. The differences in predictive performance between models with and without a DL score were statistically significant (differences in test log-likelihood 6.7 and 51.1 natural log units) but the increments in C-statistics from 0.820 to 0.822 and from 0.709 to 0.711 for T1DM and T2DM respectively, were small. CONCLUSIONS: These results show that in people with diabetes, retinal photographs contain information on future CVD risk. However for this to contribute appreciably to clinical prediction of CVD further approaches, including exploitation of serial images, need to be evaluated.

Deep learning detection of diabetic retinopathy in Scotland's diabetic eye screening programme.

BACKGROUND/AIMS: Support vector machine-based automated grading (known as iGradingM) has been shown to be safe, cost-effective and robust in the diabetic retinopathy (DR) screening (DES) programme in Scotland. It triages screening episodes as gradable with no DR versus manual grading required. The study aim was to develop a deep learning-based autograder using images and gradings from DES and to compare its performance with that of iGradingM. METHODS: Retinal images, quality assurance (QA) data and routine DR grades were obtained from national datasets in 179 944 patients for years 2006-2016. QA grades were available for 744 images. We developed a deep learning-based algorithm to detect whether either eye contained ungradable images or any DR. The sensitivity and specificity were evaluated against consensus QA grades and routine grades. RESULTS: Images used in QA which were ungradable or with DR were detected by deep learning with better specificity compared with manual graders (p<0.001) and with iGradingM (p<0.001) at the same sensitivities. Any DR according to the DES final grade was detected with 89.19% (270 392/303 154) sensitivity and 77.41% (500 945/647 158) specificity. Observable disease and referable disease were detected with sensitivities of 96.58% (16 613/17 201) and 98.48% (22 600/22 948), respectively. Overall, 43.84% of screening episodes would require manual grading. CONCLUSION: A deep learning-based system for DR grading was evaluated in QA data and images from 11 years in 50% of people attending a national DR screening programme. The system could reduce the manual grading workload at the same sensitivity compared with the current automated grading system.

Deep attention super-resolution of brain magnetic resonance images acquired under clinical protocols.

Vast quantities of Magnetic Resonance Images (MRI) are routinely acquired in clinical practice but, to speed up acquisition, these scans are typically of a quality that is sufficient for clinical diagnosis but sub-optimal for large-scale precision medicine, computational diagnostics, and large-scale neuroimaging collaborative research. Here, we present a critic-guided framework to upsample low-resolution (often 2D) MRI full scans to help overcome these limitations. We incorporate feature-importance and self-attention methods into our model to improve the interpretability of this study. We evaluate our framework on paired low- and high-resolution brain MRI structural full scans (i.e., T1-, T2-weighted, and FLAIR sequences are simultaneously input) obtained in clinical and research settings from scanners manufactured by Siemens, Phillips, and GE. We show that the upsampled MRIs are qualitatively faithful to the ground-truth high-quality scans (PSNR = 35.39; MAE = 3.78E-3; NMSE = 4.32E-10; SSIM = 0.9852; mean normal-appearing gray/white matter ratio intensity differences ranging from 0.0363 to 0.0784 for FLAIR, from 0.0010 to 0.0138 for T1-weighted and from 0.0156 to 0.074 for T2-weighted sequences). The automatic raw segmentation of tissues and lesions using the super-resolved images has fewer false positives and higher accuracy than those obtained from interpolated images in protocols represented with more than three sets in the training sample, making our approach a strong candidate for practical application in clinical and collaborative research.

Reproducibility of tract segmentation between sessions using an unsupervised modelling-based approach.

This work describes a reproducibility analysis of scalar water diffusion parameters, measured within white matter tracts segmented using a probabilistic shape modelling method. In common with previously reported neighbourhood tractography (NT) work, the technique optimises seed point placement for fibre tracking by matching the tracts generated using a number of candidate points against a reference tract, which is derived from a white matter atlas in the present study. No direct constraints are applied to the fibre tracking results. An Expectation-Maximisation algorithm is used to fully automate the procedure, and make dramatically more efficient use of data than earlier NT methods. Within-subject and between-subject variances for fractional anisotropy and mean diffusivity within the tracts are then separated using a random effects model. We find test-retest coefficients of variation (CVs) similar to those reported in another study using landmark-guided single seed points; and subject to subject CVs similar to a constraint-based multiple ROI method. We conclude that our approach is at least as effective as other methods for tract segmentation using tractography, whilst also having some additional benefits, such as its provision of a goodness-of-match measure for each segmentation.

Individualized prediction of psychosis in subjects with an at-risk mental state.

Early intervention strategies in psychosis would significantly benefit from the identification of reliable prognostic biomarkers. Pattern classification methods have shown the feasibility of an early diagnosis of psychosis onset both in clinical and familial high-risk populations. Here we were interested in replicating our previous classification findings using an independent cohort at clinical high risk for psychosis, drawn from the prospective FePsy (Fruherkennung von Psychosen) study. The same neuroanatomical-based pattern classification pipeline, consisting of a linear Support Vector Machine (SVM) and a Recursive Feature Selection (RFE) achieved 74% accuracy in predicting later onset of psychosis. The discriminative neuroanatomical pattern underlying this finding consisted of many brain areas across all four lobes and the cerebellum. These results provide proof-of-concept that the early diagnosis of psychosis is feasible using neuroanatomical-based pattern recognition.

Tract shape modelling provides evidence of topological change in corpus callosum genu during normal ageing.

Understanding how ageing affects brain structure is an important challenge for medical science. By allowing segmentation of fasciculi-of-interest from diffusion magnetic resonance imaging (dMRI) data, tractography provides a promising tool for assessing white matter connectivity in old age. However, the output from tractography algorithms is usually strongly dependent on the subjective location of user-specified seed points, with the result that it can be both difficult and time consuming to identify the same tract reliably in cross-sectional studies. Here we investigate whether a novel method for automatic single seed point placement based on tract shape modelling, termed probabilistic model-based neighbourhood tractography (PNT), can reliably segment the same tract from subject to subject in a non-demented cohort aged over 65 years. For the fasciculi investigated (genu and splenium of corpus callosum, cingulum cingulate gyri, corticospinal tracts and uncinate fasciculi), PNT was able to provide anatomically plausible representations of the tract in question in 70 to 90% of subjects compared with 2.5 to 60% if single seed points were simply transferred directly from standard to native space. In corpus callosum genu there was a significant negative correlation between a PNT-derived measure of tract shape similarity to a young brain reference tract and age, and a trend towards a significant negative correlation between tract-averaged fractional anisotropy and age; results that are consistent with previous dMRI studies of normal ageing. These data show that it is possible automatically to segment comparable tracts in the brains of older subjects using single seed point tractography, if the seed point is carefully chosen.

A probabilistic model-based approach to consistent white matter tract segmentation.

Since the invention of diffusion magnetic resonance imaging (dMRI), currently the only established method for studying white matter connectivity in a clinical environment, there has been a great deal of interest in the effects of various pathologies on the connectivity of the brain. As methods for in vivo tractography have been developed, it has become possible to track and segment specific white matter structures of interest for particular study. However, the consistency and reproducibility of tractography-based segmentation remain limited, and attempts to improve them have thus far typically involved the imposition of strong constraints on the tract reconstruction process itself. In this work we take a different approach, developing a formal probabilistic model for the relationships between comparable tracts in different scans, and then using it to choose a tract, a posteriori, which best matches a predefined reference tract for the structure of interest. We demonstrate that this method is able to significantly improve segmentation consistency without directly constraining the tractography algorithm.

Improved individualized prediction of schizophrenia in subjects at familial high risk, based on neuroanatomical data, schizotypal and neurocognitive features.

To date, there are no reliable markers for predicting onset of schizophrenia in individuals at high risk (HR). Substantial promise is, however, shown by a variety of pattern classification approaches to neuroimaging data. Here, we examined the predictive accuracy of support vector machine (SVM) in later diagnosing schizophrenia, at a single-subject level, using a cohort of HR individuals drawn from multiply affected families and a combination of neuroanatomical, schizotypal and neurocognitive variables. Baseline structural magnetic resonance imaging (MRI), schizotypal and neurocognitive data from 17 HR subjects, who subsequently developed schizophrenia and a matched group of 17 HR subjects who did not make the transition, yet had psychotic symptoms, were included in the analysis. We employed recursive feature elimination (RFE), in a nested cross-validation scheme to identify the most significant predictors of disease transition and enhance diagnostic performance. Classification accuracy was 94% when a self-completed measure of schizotypy, a declarative memory test and structural MRI data were combined into a single learning algorithm; higher than when either quantitative measure was used alone. The discriminative neuroanatomical pattern involved gray matter volume differences in frontal, orbito-frontal and occipital lobe regions bilaterally as well as parts of the superior, medial temporal lobe and cerebellar regions. Our findings suggest that an early SVM-based prediction of schizophrenia is possible and can be improved by combining schizotypal and neurocognitive features with neuroanatomical variables. However, our predictive model needs to be tested by classifying a new, independent HR cohort in order to estimate its validity.

Brain white matter structure and information processing speed in healthy older age.

Cognitive decline, especially the slowing of information processing speed, is associated with normal ageing. This decline may be due to brain cortico-cortical disconnection caused by age-related white matter deterioration. We present results from a large, narrow age range cohort of generally healthy, community-dwelling subjects in their seventies who also had their cognitive ability tested in youth (age 11 years). We investigate associations between older age brain white matter structure, several measures of information processing speed and childhood cognitive ability in 581 subjects. Analysis of diffusion tensor MRI data using Tract-based Spatial Statistics (TBSS) showed that all measures of information processing speed, as well as a general speed factor composed from these tests (g speed), were significantly associated with fractional anisotropy (FA) across the white matter skeleton rather than in specific tracts. Cognitive ability measured at age 11 years was not associated with older age white matter FA, except for the g speed-independent components of several individual processing speed tests. These results indicate that quicker and more efficient information processing requires global connectivity in older age, and that associations between white matter FA and information processing speed (both individual test scores and g speed), unlike some other aspects of later life brain structure, are generally not accounted for by cognitive ability measured in youth.

The Supervised Hierarchical Dirichlet Process.

We propose the supervised hierarchical Dirichlet process (sHDP), a nonparametric generative model for the joint distribution of a group of observations and a response variable directly associated with that whole group. We compare the sHDP with another leading method for regression on grouped data, the supervised latent Dirichlet allocation (sLDA) model. We evaluate our method on two real-world classification problems and two real-world regression problems. Bayesian nonparametric regression models based on the Dirichlet process, such as the Dirichlet process-generalised linear models (DP-GLM) have previously been explored; these models allow flexibility in modelling nonlinear relationships. However, until now, hierarchical Dirichlet process (HDP) mixtures have not seen significant use in supervised problems with grouped data since a straightforward application of the HDP on the grouped data results in learnt clusters that are not predictive of the responses. The sHDP solves this problem by allowing for clusters to be learnt jointly from the group structure and from the label assigned to each group.

Adaptive thresholding for reliable topological inference in single subject fMRI analysis.

Single subject fMRI has proved to be a useful tool for mapping functional areas in clinical procedures such as tumor resection. Using fMRI data, clinicians assess the risk, plan and execute such procedures based on thresholded statistical maps. However, because current thresholding methods were developed mainly in the context of cognitive neuroscience group studies, most single subject fMRI maps are thresholded manually to satisfy specific criteria related to single subject analyzes. Here, we propose a new adaptive thresholding method which combines Gamma-Gaussian mixture modeling with topological thresholding to improve cluster delineation. In a series of simulations we show that by adapting to the signal and noise properties, the new method performs well in terms of total number of errors but also in terms of the trade-off between false negative and positive cluster error rates. Similarly, simulations show that adaptive thresholding performs better than fixed thresholding in terms of over and underestimation of the true activation border (i.e., higher spatial accuracy). Finally, through simulations and a motor test-retest study on 10 volunteer subjects, we show that adaptive thresholding improves reliability, mainly by accounting for the global signal variance. This in turn increases the likelihood that the true activation pattern can be determined offering an automatic yet flexible way to threshold single subject fMRI maps.

Improved segmentation reproducibility in group tractography using a quantitative tract similarity measure.

The field of tractography is rapidly developing, and many automatic or semiautomatic algorithms have now been devised to segment and visualize neural white matter fasciculi in vivo. However, these algorithms typically need to be given a starting location as input, and their output can be strongly dependent on the exact location of this "seed point". No robust method has yet been devised for placing these seed points so as to segment a comparable tract in a group of subjects. Here, we develop a measure of tract similarity, based on the shapes and lengths of the two tracts being compared, and apply it to the problem of consistent seed point placement and tract segmentation in group data. We demonstrate that using a single seed point transferred from standard space to each native space produces considerable variability in tractography output between scans. However, by seeding in a group of nearby candidate points and choosing the output with the greatest similarity to a reference tract chosen in advance--a method we refer to as neighborhood tractography--this variability can be significantly reduced.