Hyperparameter selection for dataset-constrained semantic segmentation: Practical machine learning optimization.

PURPOSE/AIM: This paper provides a pedagogical example for systematic machine learning optimization in small dataset image segmentation, emphasizing hyperparameter selections. A simple process is presented for medical physicists to examine hyperparameter optimization. This is also applied to a case-study, demonstrating the benefit of the method. MATERIALS AND METHODS: An unrestricted public Computed Tomography (CT) dataset, with binary organ segmentation, was used to develop a multiclass segmentation model. To start the optimization process, a preliminary manual search of hyperparameters was conducted and from there a grid search identified the most influential result metrics. A total of 658 different models were trained in 2100 h, using 13 160 effective patients. The quantity of results was analyzed using random forest regression, identifying relative hyperparameter impact. RESULTS: Metric implied segmentation quality (accuracy 96.8%, precision 95.1%) and visual inspection were found to be mismatched. In this work batch normalization was most important, but performance varied with hyperparameters and metrics selected. Targeted grid-search optimization and random forest analysis of relative hyperparameter importance, was an easily implementable sensitivity analysis approach. CONCLUSION: The proposed optimization method gives a systematic and quantitative approach to something intuitively understood, that hyperparameters change model performance. Even just grid search optimization with random forest analysis presented here can be informative within hardware and data quality/availability limitations, adding confidence to model validity and minimize decision-making risks. By providing a guided methodology, this work helps medical physicists to improve their model optimization, irrespective of specific challenges posed by datasets and model design.

Multimodal nonlinear correlates of behavioural symptoms in frontotemporal dementia.

Studies exploring the brain correlates of behavioral symptoms in the frontotemporal dementia spectrum (FTD) have mainly searched for linear correlations with single modality neuroimaging data, either structural magnetic resonance imaging (MRI) or fluoro-deoxy-D-glucose positron emission tomography (FDG-PET). We aimed at studying the two imaging modalities in combination to identify nonlinear co-occurring patterns of atrophy and hypometabolism related to behavioral symptoms. We analyzed data from 93 FTD patients who underwent T1-weighted MRI, FDG-PET imaging, and neuropsychological assessment including the Neuropsychiatric Inventory, Frontal Systems Behavior Scale, and Neurobehavioral Rating Scale. We used a data-driven approach to identify the principal components underlying behavioral variability, then related the identified components to brain variability using a newly developed method fusing maps of grey matter volume and FDG metabolism. A component representing apathy, executive dysfunction, and emotional withdrawal was associated with atrophy in bilateral anterior insula and putamen, and with hypometabolism in the right prefrontal cortex. Another component representing the disinhibition versus depression/mutism continuum was associated with atrophy in the right striatum and ventromedial prefrontal cortex for disinhibition, and hypometabolism in the left fronto-opercular region and sensorimotor cortices for depression/mutism. A component representing psychosis was associated with hypometabolism in the prefrontal cortex and hypermetabolism in auditory and visual cortices. Behavioral symptoms in FTD are associated with atrophy and altered metabolism of specific brain regions, especially located in the frontal lobes, in a hierarchical way: apathy and disinhibition are mostly associated with grey matter atrophy, whereas psychotic symptoms are mostly associated with hyper-/hypo-metabolism.

niimath and fslmaths: replication as a method to enhance popular neuroimaging tools.

Neuroimaging involves the acquisition of extensive 3D images and 4D time series data to gain insights into brain structure and function. The analysis of such data necessitates both spatial and temporal processing. In this context, "fslmaths" has established itself as a foundational software tool within our field, facilitating domain-specific image processing. Here, we introduce "niimath," a clone of fslmaths. While the term "clone" often carries negative connotations, we illustrate the merits of replicating widely-used tools, touching on aspects of licensing, performance optimization, and portability. For instance, our work enables the popular functions of fslmaths to be disseminated in various forms, such as a high-performance compiled R package known as "imbibe", a Windows executable, and a WebAssembly plugin compatible with JavaScript. This versatility is demonstrated through our NiiVue live demo web page. This application allows 'edge computing' where image processing can be done with a zero-footprint tool that runs on any web device without requiring private data to be shared to the cloud. Furthermore, our efforts have contributed back to FSL, which has integrated the optimizations that we've developed. This synergy has enhanced the overall transparency, utility and efficiency of tools widely relied upon in the neuroimaging community.

Anatomically plausible segmentations: Explicitly preserving topology through prior deformations.

Since the rise of deep learning, new medical segmentation methods have rapidly been proposed with extremely promising results, often reporting marginal improvements on the previous state-of-the-art (SOTA) method. However, on visual inspection errors are often revealed, such as topological mistakes (e.g. holes or folds), that are not detected using traditional evaluation metrics. Incorrect topology can often lead to errors in clinically required downstream image processing tasks. Therefore, there is a need for new methods to focus on ensuring segmentations are topologically correct. In this work, we present TEDS-Net: a segmentation network that preserves anatomical topology whilst maintaining segmentation performance that is competitive with SOTA baselines. Further, we show how current SOTA segmentation methods can introduce problematic topological errors. TEDS-Net achieves anatomically plausible segmentation by using learnt topology-preserving fields to deform a prior. Traditionally, topology-preserving fields are described in the continuous domain and begin to break down when working in the discrete domain. Here, we introduce additional modifications that more strictly enforce topology preservation. We illustrate our method on an open-source medical heart dataset, performing both single and multi-structure segmentation, and show that the generated fields contain no folding voxels, which corresponds to full topology preservation on individual structures whilst vastly outperforming the other baselines on overall scene topology. The code is available at: https://github.com/mwyburd/TEDS-Net.

Predictions for functional outcome and mortality in acute ischaemic stroke following successful endovascular thrombectomy.

Background: Accurate outcome predictions for patients who had ischaemic stroke with successful reperfusion after endovascular thrombectomy (EVT) may improve patient treatment and care. Our study developed prediction models for key clinical outcomes in patients with successful reperfusion following EVT in an Australian population. Methods: The study included all patients who had ischaemic stroke with occlusion in the proximal anterior cerebral circulation and successful reperfusion post-EVT over a 7-year period. Multivariable logistic regression and Cox regression models, incorporating bootstrap and multiple imputation techniques, were used to identify predictors and develop models for key clinical outcomes: 3-month poor functional status; 30-day, 1-year and 3-year mortality; survival time. Results: A total of 978 patients were included in the analyses. Predictors associated with one or more poor outcomes include: older age (ORs for every 5-year increase: 1.22-1.40), higher premorbid functional modified Rankin Scale (ORs: 1.31-1.75), higher baseline National Institutes of Health Stroke Scale (ORs: 1.05-1.07) score, higher blood glucose (ORs: 1.08-1.19), larger core volume (ORs for every 10 mL increase: 1.10-1.22), pre-EVT thrombolytic therapy (ORs: 0.44-0.56), history of heart failure (outcome: 30-day mortality, OR=1.87), interhospital transfer (ORs: 1.42 to 1.53), non-rural/regional stroke onset (outcome: functional dependency, OR=0.64), longer onset-to-groin puncture time (outcome: 3-year mortality, OR=1.08) and atherosclerosis-caused stroke (outcome: functional dependency, OR=1.68). The models using these predictors demonstrated moderate predictive abilities (area under the receiver operating characteristic curve range: 0.752-0.796). Conclusion: Our models using real-world predictors assessed at hospital admission showed satisfactory performance in predicting poor functional outcomes and short-term and long-term mortality for patients with successful reperfusion following EVT. These can be used to inform EVT treatment provision and consent.

The prevalence and topography of spinal cord demyelination in multiple sclerosis: a retrospective study.

Spinal cord pathology is a major determinant of irreversible disability in progressive multiple sclerosis. The demyelinated lesion is a cardinal feature. The well-characterised anatomy of the spinal cord and new analytic approaches allows the systematic study of lesion topography and its extent of inflammatory activity unveiling new insights into disease pathogenesis. We studied cervical, thoracic, and lumbar spinal cord tissue from 119 pathologically confirmed multiple sclerosis cases. Immunohistochemistry was used to detect demyelination (PLP) and classify lesional inflammatory activity (CD68). Prevalence and distribution of demyelination, staged by lesion activity, was determined and topographical maps were created to identify patterns of lesion prevalence and distribution using mixed models and permutation-based voxelwise analysis. 460 lesions were observed throughout the spinal cord with 76.5% of cases demonstrating at least 1 lesion. The cervical level was preferentially affected by lesions. 58.3% of lesions were inflammatory with 87.9% of cases harbouring at least 1 inflammatory lesion. Topographically, lesions consistently affected the dorsal and lateral columns with relative sparing of subpial areas in a distribution mirroring the vascular network. The presence of spinal cord lesions and the proportion of active lesions related strongly with clinical disease milestones, including time from onset to wheelchair and onset to death. We demonstrate that spinal cord demyelination is common, highly inflammatory, has a predilection for the cervical level, and relates to clinical disability. The topography of lesions in the dorsal and lateral columns and relative sparing of subpial areas points to a role of the vasculature in lesion pathogenesis, suggesting short-range cell infiltration from the blood and signaling molecules circulating in the perivascular space incite lesion development. These findings challenge the notion that end-stage progressive multiple sclerosis is 'burnt out' and an outside-in lesional gradient predominates in the spinal cord. Taken together, this study provides support for long-term targeting of inflammatory demyelination in the spinal cord and nominates vascular dysfunction as a potential target for new therapeutic approaches to limit irreversible disability.

Cardiac function may be compromised in patients with elevated blood cobalt levels secondary to metal-on-metal hip implants.

Aims: Elevated blood cobalt levels secondary to metal-on-metal (MoM) hip arthroplasties are a suggested risk factor for developing cardiovascular complications including cardiomyopathy. Clinical studies assessing patients with MoM hips using left ventricular ejection fraction (LVEF) have found conflicting evidence of cobalt-induced cardiomyopathy. Global longitudinal strain (GLS) is an echocardiography measurement known to be more sensitive than LVEF when diagnosing early cardiomyopathies. The extent of cardiovascular injury, as measured by GLS, in patients with elevated blood cobalt levels has not previously been examined. Methods: A total of 16 patients with documented blood cobalt ion levels above 13 µg/l (13 ppb, 221 nmol/l) were identified from a regional arthroplasty database. They were matched with eight patients awaiting hip arthroplasty. All patients underwent echocardiography, including GLS, investigating potential signs of cardiomyopathy. Results: Patients with MoM hip arthroplasties had a mean blood cobalt level of 29 µg/l (495 nmol/l) compared to 0.01 µg/l (0.2 nmol/l) in the control group. GLS readings were available for seven of the MoM cohort, and were significantly lower when compared with controls (-15.5% vs -18% (MoM vs control); p = 0.025)). Pearson correlation demonstrated that GLS significantly correlated with blood cobalt level (r = 0.8521; p < 0.001). However, there were no differences or correlations for other echocardiography measurements, including LVEF (64.3% vs 63.7% (MoM vs control); p = 0.845). Conclusion: This study supports the hypothesis that patients with elevated blood cobalt levels above 13 µg/l in the presence of a MoM hip implant may have impaired cardiac function compared to a control group of patients awaiting hip arthroplasty. It is the first study to use the more sensitive parameter of GLS to assess for any cardiac contractile dysfunction in patients with a MoM hip implant and a normal LVEF. Larger studies should be performed to determine the potential of GLS as a predictor of cardiac complications in patients with MoM arthroplasties.

Acetabular retroversion: functional or anatomical?

Aims: The aim of this study is to evaluate whether acetabular retroversion (AR) represents a structural anatomical abnormality of the pelvis or is a functional phenomenon of pelvic positioning in the sagittal plane, and to what extent the changes that result from patient-specific functional position affect the extent of AR. Methods: A comparative radiological study of 19 patients (38 hips) with AR were compared with a control group of 30 asymptomatic patients (60 hips). CT scans were corrected for rotation in the axial and coronal planes, and the sagittal plane was then aligned to the anterior pelvic plane. External rotation of the hemipelvis was assessed using the superior iliac wing and inferior iliac wing angles as well as quadrilateral plate angles, and correlated with cranial and central acetabular version. Sagittal anatomical parameters were also measured and correlated to version measurements. In 12 AR patients (24 hips), the axial measurements were repeated after matching sagittal pelvic rotation with standing and supine anteroposterior radiographs. Results: Acetabular version was significantly lower and measurements of external rotation of the hemipelvis were significantly increased in the AR group compared to the control group. The AR group also had increased evidence of anterior projection of the iliac wing in the sagittal plane. The acetabular orientation angles were more retroverted in the supine compared to standing position, and the change in acetabular version correlated with the change in sagittal pelvic tilt. An anterior pelvic tilt of 1° correlated with 1.02° of increased cranial retroversion and 0.76° of increased central retroversion. Conclusion: This study has demonstrated that patients with symptomatic AR have both an externally rotated hemipelvis and increased anterior projection of the iliac wing compared to a control group of asymptomatic patients. Functional sagittal pelvic positioning was also found to affect AR in symptomatic patients: the acetabulum was more retroverted in the supine position compared to standing position. Changes in acetabular version correlate with the change in sagittal pelvic tilt. These findings should be taken into account by surgeons when planning acetabular correction for AR with periacetabular osteotomy.

Subthalamic nucleus shows opposite functional connectivity pattern in Huntington's and Parkinson's disease.

Huntington's and Parkinson's disease are two movement disorders representing mainly opposite states of the basal ganglia inhibitory function. Despite being an integral part of the cortico-subcortico-cortical circuitry, the subthalamic nucleus function has been studied at the level of detail required to isolate its signal only through invasive studies in Huntington's and Parkinson's disease. Here, we tested whether the subthalamic nucleus exhibited opposite functional signatures in early Huntington's and Parkinson's disease. We included both movement disorders in the same whole-brain imaging study, and leveraged ultra-high-field 7T MRI to achieve the very fine resolution needed to investigate the smallest of the basal ganglia nuclei. Eleven of the 12 Huntington's disease carriers were recruited at a premanifest stage, while 16 of the 18 Parkinson's disease patients only exhibited unilateral motor symptoms (15 were at Stage I of Hoehn and Yahr off medication). Our group comparison interaction analyses, including 24 healthy controls, revealed a differential effect of Huntington's and Parkinson's disease on the functional connectivity at rest of the subthalamic nucleus within the sensorimotor network, i.e. an opposite effect compared with their respective age-matched healthy control groups. This differential impact in the subthalamic nucleus included an area precisely corresponding to the deep brain stimulation 'sweet spot'-the area with maximum overall efficacy-in Parkinson's disease. Importantly, the severity of deviation away from controls' resting-state values in the subthalamic nucleus was associated with the severity of motor and cognitive symptoms in both diseases, despite functional connectivity going in opposite directions in each disorder. We also observed an altered, opposite impact of Huntington's and Parkinson's disease on functional connectivity within the sensorimotor cortex, once again with relevant associations with clinical symptoms. The high resolution offered by the 7T scanner has thus made it possible to explore the complex interplay between the disease effects and their contribution on the subthalamic nucleus, and sensorimotor cortex. Taken altogether, these findings reveal for the first time non-invasively in humans a differential, clinically meaningful impact of the pathophysiological process of these two movement disorders on the overall sensorimotor functional connection of the subthalamic nucleus and sensorimotor cortex.

Normative spatiotemporal fetal brain maturation with satisfactory development at 2 years.

Maturation of the human fetal brain should follow precisely scheduled structural growth and folding of the cerebral cortex for optimal postnatal function1. We present a normative digital atlas of fetal brain maturation based on a prospective international cohort of healthy pregnant women2, selected using World Health Organization recommendations for growth standards3. Their fetuses were accurately dated in the first trimester, with satisfactory growth and neurodevelopment from early pregnancy to 2 years of age4,5. The atlas was produced using 1,059 optimal quality, three-dimensional ultrasound brain volumes from 899 of the fetuses and an automated analysis pipeline6-8. The atlas corresponds structurally to published magnetic resonance images9, but with finer anatomical details in deep grey matter. The between-study site variability represented less than 8.0% of the total variance of all brain measures, supporting pooling data from the eight study sites to produce patterns of normative maturation. We have thereby generated an average representation of each cerebral hemisphere between 14 and 31 weeks' gestation with quantification of intracranial volume variability and growth patterns. Emergent asymmetries were detectable from as early as 14 weeks, with peak asymmetries in regions associated with language development and functional lateralization between 20 and 26 weeks' gestation. These patterns were validated in 1,487 three-dimensional brain volumes from 1,295 different fetuses in the same cohort. We provide a unique spatiotemporal benchmark of fetal brain maturation from a large cohort with normative postnatal growth and neurodevelopment.

Multimodal nonlinear correlates of behavioural symptoms in frontotemporal dementia.

Background: Studies exploring the brain correlates of behavioural symptoms in the frontotemporal dementia spectrum (FTD) have mainly searched for linear correlations with single modality neuroimaging data, either structural magnetic resonance imaging (MRI) or fluoro-deoxy-D-glucose positron emission tomography (FDG-PET). We aimed at studying the two imaging modalities in combination to identify nonlinear co-occurring patterns of atrophy and hypometabolism related to behavioural symptoms. Methods: We analysed data from 93 FTD patients who underwent T1-weighted MRI, FDG-PET imaging, and neuropsychological assessment including the Neuropsychiatric Inventory, Frontal Systems Behaviour Scale, and Neurobehavioral Rating Scale. We used a data-driven approach to identify the principal components underlying behavioural variability, then related the identified components to brain variability using a newly developed method fusing maps of grey matter volume and FDG metabolism. Results: A component representing apathy, executive dysfunction, and emotional withdrawal was associated with atrophy in bilateral anterior insula and putamen, and with hypometabolism in the right prefrontal cortex. Another component representing the disinhibition versus depression/mutism continuum was associated with atrophy in the right striatum and ventromedial prefrontal cortex for disinhibition, and hypometabolism in the left fronto-opercular region and sensorimotor cortices for depression/mutism. A component representing psychosis was associated with hypometabolism in the prefrontal cortex and hypermetabolism in auditory and visual cortices. Discussion: Behavioural symptoms in FTD are associated with atrophy and altered metabolism of specific brain regions, especially located in the frontal lobes, in a hierarchical way: apathy and disinhibition are mostly associated with grey matter atrophy, whereas psychotic symptoms are mostly associated with hyper-/hypo-metabolism.

BIANCA-MS: An optimized tool for automated multiple sclerosis lesion segmentation.

In this work we present BIANCA-MS, a novel tool for brain white matter lesion segmentation in multiple sclerosis (MS), able to generalize across both the wide spectrum of MRI acquisition protocols and the heterogeneity of manually labeled data. BIANCA-MS is based on the original version of BIANCA and implements two innovative elements: a harmonized setting, tested under different MRI protocols, which avoids the need to further tune algorithm parameters to each dataset; and a cleaning step developed to improve consistency in automated and manual segmentations, thus reducing unwanted variability in output segmentations and validation data. BIANCA-MS was tested on three datasets, acquired with different MRI protocols. First, we compared BIANCA-MS to other widely used tools. Second, we tested how BIANCA-MS performs in separate datasets. Finally, we evaluated BIANCA-MS performance on a pooled dataset where all MRI data were merged. We calculated the overlap using the DICE spatial similarity index (SI) as well as the number of false positive/negative clusters (nFPC/nFNC) in comparison to the manual masks processed with the cleaning step. BIANCA-MS clearly outperformed other available tools in both high- and low-resolution images and provided comparable performance across different scanning protocols, sets of modalities and image resolutions. BIANCA-MS performance on the pooled dataset (SI: 0.72 ± 0.25, nFPC: 13 ± 11, nFNC: 4 ± 8) were comparable to those achieved on each individual dataset (median across datasets SI: 0.72 ± 0.28, nFPC: 14 ± 11, nFNC: 4 ± 8). Our findings suggest that BIANCA-MS is a robust and accurate approach for automated MS lesion segmentation.

Automated detection of cerebral microbleeds on MR images using knowledge distillation framework.

Introduction: Cerebral microbleeds (CMBs) are associated with white matter damage, and various neurodegenerative and cerebrovascular diseases. CMBs occur as small, circular hypointense lesions on T2*-weighted gradient recalled echo (GRE) and susceptibility-weighted imaging (SWI) images, and hyperintense on quantitative susceptibility mapping (QSM) images due to their paramagnetic nature. Accurate automated detection of CMBs would help to determine quantitative imaging biomarkers (e.g., CMB count) on large datasets. In this work, we propose a fully automated, deep learning-based, 3-step algorithm, using structural and anatomical properties of CMBs from any single input image modality (e.g., GRE/SWI/QSM) for their accurate detections. Methods: In our method, the first step consists of an initial candidate detection step that detects CMBs with high sensitivity. In the second step, candidate discrimination step is performed using a knowledge distillation framework, with a multi-tasking teacher network that guides the student network to classify CMB and non-CMB instances in an offline manner. Finally, a morphological clean-up step further reduces false positives using anatomical constraints. We used four datasets consisting of different modalities specified above, acquired using various protocols and with a variety of pathological and demographic characteristics. Results: On cross-validation within datasets, our method achieved a cluster-wise true positive rate (TPR) of over 90% with an average of <2 false positives per subject. The knowledge distillation framework improves the cluster-wise TPR of the student model by 15%. Our method is flexible in terms of the input modality and provides comparable cluster-wise TPR and better cluster-wise precision compared to existing state-of-the-art methods. When evaluating across different datasets, our method showed good generalizability with a cluster-wise TPR >80 % with different modalities. The python implementation of the proposed method is openly available.

Differential Structural Brain Changes Between Responders and Nonresponders After Physical Exercise Therapy for Chronic Nonspecific Neck Pain.

OBJECTIVES: Physical exercise therapy is effective for some people with chronic nonspecific neck pain but not for others. Differences in exercise-induced pain-modulatory responses are likely driven by brain changes. We investigated structural brain differences at baseline and changes after an exercise intervention. The primary aim was to investigate changes in structural brain characteristics after physical exercise therapy for people with chronic nonspecific neck pain. The secondary aims were to investigate (1) baseline differences in structural brain characteristics between responders and nonresponders to exercise therapy, and (2) differential brain changes after exercise therapy between responders and nonresponders. MATERIALS AND METHODS: This was a prospective longitudinal cohort study. Twenty-four participants (18 females, mean age 39.7 y) with chronic nonspecific neck pain were included. Responders were selected as those with ≥20% improvement in Neck Disability Index. Structural magnetic resonance imaging was obtained before and after an 8-week physical exercise intervention delivered by a physiotherapist. Freesurfer cluster-wise analyses were performed and supplemented with an analysis of pain-specific brain regions of interest. RESULTS: Various changes in grey matter volume and thickness were found after the intervention, for example, frontal cortex volume decreased (cluster-weighted P value = 0.0002, 95% CI: 0.0000-0.0004). We found numerous differences between responders and nonresponders, most notably, after the exercise intervention bilateral insular volume decreased in responders, but increased in nonresponders (cluster-weighted P value ≤ 0.0002). DISCUSSION: The brain changes found in this study may underpin clinically observed differential effects between responders and nonresponders to exercise therapy for people with chronic neck pain. Identification of these changes is an important step toward personalized treatment approaches.

A comparison of lesion mapping analyses based on CT versus MR imaging in stroke.

It is commonly asserted that MRI-derived lesion masks outperform CT-derived lesion masks in lesion-mapping analysis. However, no quantitative analysis has been conducted to support or refute this claim. This study reports an objective comparison of lesion-mapping analyses based on CT- and MRI-derived lesion masks to clarify how input imaging type may ultimately impact analysis results. Routine CT and MRI data were collected from 85 acute stroke survivors. These data were employed to create binarized lesion masks and conduct lesion-mapping analyses on simulated behavioral data. Following standard lesion-mapping analysis methodology, each voxel or region of interest (ROI) were considered as the underlying "target" within CT and MRI data independently. The resulting thresholded z-maps were compared between matched CT- and MRI-based analyses. Paired MRI- and CT-derived lesion masks were found to exhibit significant variance in location, overlap, and size. In ROI-level simulations, both CT and MRI-derived analyses yielded low Dice similarity coefficients, but CT analyses yielded a significantly higher proportion of results which overlapped with target ROIs. In single-voxel simulations, MRI-based lesion mapping was able to include more voxels than CT-based analyses, but CT-based analysis results were closer to the underlying target voxel. Simulated lesion-symptom mapping results yielded by paired CT and MRI lesion-symptom mapping analyses demonstrated moderate agreement in terms of Dice coefficient when systematic differences in cluster size and lesion overlay are considered. Overall, these results suggest that CT and MR-derived lesion-symptom mapping results do not reliably differ in accuracy. This finding is critically important as it suggests that future studies can employ CT-derived lesion masks if these scans are available within the appropriate time-window.

Normative models for neuroimaging markers: Impact of model selection, sample size and evaluation criteria.

Modelling population reference curves or normative modelling is increasingly used with the advent of large neuroimaging studies. In this paper we assess the performance of fitting methods from the perspective of clinical applications and investigate the influence of the sample size. Further, we evaluate linear and non-linear models for percentile curve estimation and highlight how the bias-variance trade-off manifests in typical neuroimaging data. We created plausible ground truth distributions of hippocampal volumes in the age range of 45 to 80 years, as an example application. Based on these distributions we repeatedly simulated samples for sizes between 50 and 50,000 data points, and for each simulated sample we fitted a range of normative models. We compared the fitted models and their variability across repetitions to the ground truth, with specific focus on the outer percentiles (1st, 5th, 10th) as these are the most clinically relevant. Our results quantify the expected decreasing trend in variance of the volume estimates with increasing sample size. However, bias in the volume estimates only decreases a modest amount, without much improvement at large sample sizes. The uncertainty of model performance is substantial for what would often be considered large samples in a neuroimaging context and rises dramatically at the ends of the age range, where fewer data points exist. Flexible models perform better across sample sizes, especially for non-linear ground truth. Surprisingly large samples of several thousand data points are needed to accurately capture outlying percentiles across the age range for applications in research and clinical settings. Performance evaluation methods should assess both bias and variance. Furthermore, caution is needed when attempting to go near the ends of the age range captured by the source data set and, as is a well known general principle, extrapolation beyond the age range should always be avoided. To help with such evaluations of normative models we have made our code available to guide researchers developing or utilising normative models.

Tensor image registration library: Deformable registration of stand-alone histology images to whole-brain post-mortem MRI data.

BACKGROUND: Accurate registration between microscopy and MRI data is necessary for validating imaging biomarkers against neuropathology, and to disentangle complex signal dependencies in microstructural MRI. Existing registration methods often rely on serial histological sampling or significant manual input, providing limited scope to work with a large number of stand-alone histology sections. Here we present a customisable pipeline to assist the registration of stand-alone histology sections to whole-brain MRI data. METHODS: Our pipeline registers stained histology sections to whole-brain post-mortem MRI in 4 stages, with the help of two photographic intermediaries: a block face image (to undistort histology sections) and coronal brain slab photographs (to insert them into MRI space). Each registration stage is implemented as a configurable stand-alone Python script using our novel platform, Tensor Image Registration Library (TIRL), which provides flexibility for wider adaptation. We report our experience of registering 87 PLP-stained histology sections from 14 subjects and perform various experiments to assess the accuracy and robustness of each stage of the pipeline. RESULTS: All 87 histology sections were successfully registered to MRI. Histology-to-block registration (Stage 1) achieved 0.2-0.4 mm accuracy, better than commonly used existing methods. Block-to-slice matching (Stage 2) showed great robustness in automatically identifying and inserting small tissue blocks into whole brain slices with 0.2 mm accuracy. Simulations demonstrated sub-voxel level accuracy (0.13 mm) of the slice-to-volume registration (Stage 3) algorithm, which was observed in over 200 actual brain slice registrations, compensating 3D slice deformations up to 6.5 mm. Stage 4 combined the previous stages and generated refined pixelwise aligned multi-modal histology-MRI stacks. CONCLUSIONS: Our open-source pipeline provides robust automation tools for registering stand-alone histology sections to MRI data with sub-voxel level precision, and the underlying framework makes it readily adaptable to a diverse range of microscopy-MRI studies.

Challenges for machine learning in clinical translation of big data imaging studies.

Combining deep learning image analysis methods and large-scale imaging datasets offers many opportunities to neuroscience imaging and epidemiology. However, despite these opportunities and the success of deep learning when applied to a range of neuroimaging tasks and domains, significant barriers continue to limit the impact of large-scale datasets and analysis tools. Here, we examine the main challenges and the approaches that have been explored to overcome them. We focus on issues relating to data availability, interpretability, evaluation, and logistical challenges and discuss the problems that still need to be tackled to enable the success of "big data" deep learning approaches beyond research.

Pre-thrombectomy prognostic prediction of large-vessel ischemic stroke using machine learning: A systematic review and meta-analysis.

Introduction: Machine learning (ML) methods are being increasingly applied to prognostic prediction for stroke patients with large vessel occlusion (LVO) treated with endovascular thrombectomy. This systematic review aims to summarize ML-based pre-thrombectomy prognostic models for LVO stroke and identify key research gaps. Methods: Literature searches were performed in Embase, PubMed, Web of Science, and Scopus. Meta-analyses of the area under the receiver operating characteristic curves (AUCs) of ML models were conducted to synthesize model performance. Results: Sixteen studies describing 19 models were eligible. The predicted outcomes include functional outcome at 90 days, successful reperfusion, and hemorrhagic transformation. Functional outcome was analyzed by 10 conventional ML models (pooled AUC=0.81, 95% confidence interval [CI]: 0.77-0.85, AUC range: 0.68-0.93) and four deep learning (DL) models (pooled AUC=0.75, 95% CI: 0.70-0.81, AUC range: 0.71-0.81). Successful reperfusion was analyzed by three conventional ML models (pooled AUC=0.72, 95% CI: 0.56-0.88, AUC range: 0.55-0.88) and one DL model (AUC=0.65, 95% CI: 0.62-0.68). Conclusions: Conventional ML and DL models have shown variable performance in predicting post-treatment outcomes of LVO without generally demonstrating superiority compared to existing prognostic scores. Most models were developed using small datasets, lacked solid external validation, and at high risk of potential bias. There is considerable scope to improve study design and model performance. The application of ML and DL methods to improve the prediction of prognosis in LVO stroke, while promising, remains nascent. Systematic review registration: https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42021266524, identifier CRD42021266524.

Mutual Information Neural Estimation for Unsupervised Multi-Modal Registration of Brain Images.

Many applications in image-guided surgery and therapy require fast and reliable non-linear, multi-modal image registration. Recently proposed unsupervised deep learning-based registration methods have demonstrated superior per-formance compared to iterative methods in just a fraction of the time. Most of the learning-based methods have focused on mono-modal image registration. The extension to multi-modal registration depends on the use of an appropriate similarity function, such as the mutual information (MI). We propose guiding the training of a deep learning-based registration method with MI estimation between an image-pair in an end-to-end trainable network. Our results show that a small, 2-layer network produces competitive results in both mono- and multi-modal registration, with sub-second run-times. Comparisons to both iterative and deep learning-based methods show that our MI-based method produces topologically and qualitatively superior results with an extremely low rate of non-diffeomorphic transformations. Real-time clinical application will benefit from a better visual matching of anatomical structures and less registration failures/outliers.