White matter hyperintensities segmentation using an ensemble of neural networks

A 25-Year Retrospective of the Use of AI for Diagnosing Acute Stroke: Systematic Review

BACKGROUND

Stroke is a leading cause of death and disability worldwide. Rapid and accurate diagnosis is crucial for minimizing brain damage and optimizing treatment plans.

OBJECTIVE

This review aims to summarize the methods of artificial intelligence (AI)-assisted stroke diagnosis over the past 25 years, providing an overview of performance metrics and algorithm development trends. It also delves into existing issues and future prospects, intending to offer a comprehensive reference for clinical practice.

METHODS

A total of 50 representative articles published between 1999 and 2024 on using AI technology for stroke prevention and diagnosis were systematically selected and analyzed in detail.

RESULTS

AI-assisted stroke diagnosis has made significant advances in stroke lesion segmentation and classification, stroke risk prediction, and stroke prognosis. Before 2012, research mainly focused on segmentation using traditional thresholding and heuristic techniques. From 2012 to 2016, the focus shifted to machine learning (ML)-based approaches. After 2016, the emphasis moved to deep learning (DL), which brought significant improvements in accuracy. In stroke lesion segmentation and classification as well as stroke risk prediction, DL has shown superiority over ML. In stroke prognosis, both DL and ML have shown good performance.

CONCLUSIONS

Over the past 25 years, AI technology has shown promising performance in stroke diagnosis.

Development and validation of a two-stage convolutional neural network algorithm for segmentation of MRI white matter hyperintensities for longitudinal studies in CADASIL

BACKGROUND

Segmentation of white matter hyperintensities (WMH) in CADASIL, one of the most severe cerebral small vessel disease of genetic origin, is challenging.

METHOD

We adapted and validated an automatic method based on a convolutional neural network (CNN) algorithm and using a large dataset of 2D and/or 3D FLAIR and T1-weighted images acquired in 132 patients, to measure the progression of WMH in this condition.

RESULTS

The volume of WMH measured using this method correlated strongly with reference data validated by experts. WMH segmentation was also clearly improved compared to the BIANCA segmentation method. Combining two successive learning models was found to be of particular interest, reducing the number of false-positive voxels and the extent of under-segmentation detected after a single-stage process. With the two-stage approach, WMH progression correlated with measures derived from the reference masks for lesions increasing with age, and with the variable WMH progression trajectories at individual level. We also confirmed the expected effect of the initial load of WMH and the influence of the type of MRI acquisition on measures of this progression.

CONCLUSION

Altogether, our findings suggest that WMH progression in CADASIL can be measured automatically with adequate confidence by a CNN segmentation algorithm.

Obesity and brain volumes: mediation by cardiometabolic and inflammatory measures

BACKGROUND

This study aimed to investigate the relationship between overall obesity, central obesity and brain volumes, as well as to determine the extent to which cardiometabolic and inflammatory measures act as mediators in the association between body mass index (BMI), waist-hip ratio (WHR) and brain volumes.

METHODS

In the context of counterfactual framework, mediation analysis was used to explore the potential mediation in which cardiometabolic and inflammatory measures may mediate the relationship between BMI, WHR, and brain volumes.

RESULTS

Among 2413 community-dwelling participants, those with high BMI or WHR levels experienced an approximately brain ageing of 4 years. Especially, individuals with high WHR or BMI under the age of 65 exhibited white matter hyperintensity volume (WMHV) differences equivalent to around 5 years of ageing. Conversely, in the high-level WHR population over the age of 65, premature brain ageing in gray matter volume (GMV) exceeded 4.5 years. For GMV, more than 45% of the observed effect of WHR was mediated by glycaemic metabolism indicators. This proportion increases to 78.70% when blood pressure, triglyceride, leucocyte count, and neutrophil count are jointly considered with glycaemic metabolism indicators. Regarding WHR and BMI's association with WMHV, cardiometabolic and inflammatory indicators, along with high-density lipoprotein cholesterol, mediated 35.50% and 20.20% of the respective effects.

CONCLUSIONS

Overall obesity and central obesity were associated with lower GMV and higher WMHV, a process that is partially mediated by the presence of cardiometabolic and inflammatory measures.

Evolution of white matter hyperintensity segmentation methods and implementation over the past two decades; an incomplete shift towards deep learning

This systematic review examines the prevalence, underlying mechanisms, cohort characteristics, evaluation criteria, and cohort types in white matter hyperintensity (WMH) pipeline and implementation literature spanning the last two decades. Following Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines, we categorized WMH segmentation tools based on their methodologies from January 1, 2000, to November 18, 2022. Inclusion criteria involved articles using openly available techniques with detailed descriptions, focusing on WMH as a primary outcome. Our analysis identified 1007 visual rating scales, 118 pipeline development articles, and 509 implementation articles. These studies predominantly explored aging, dementia, psychiatric disorders, and small vessel disease, with aging and dementia being the most prevalent cohorts. Deep learning emerged as the most frequently developed segmentation technique, indicative of a heightened scrutiny in new technique development over the past two decades. We illustrate observed patterns and discrepancies between published and implemented WMH techniques. Despite increasingly sophisticated quantitative segmentation options, visual rating scales persist, with the SPM technique being the most utilized among quantitative methods and potentially serving as a reference standard for newer techniques. Our findings highlight the need for future standards in WMH segmentation, and we provide recommendations based on these observations.

LST-AI: A deep learning ensemble for accurate MS lesion segmentation

Automated segmentation of brain white matter lesions is crucial for both clinical assessment and scientific research in multiple sclerosis (MS). Over a decade ago, we introduced an engineered lesion segmentation tool, LST. While recent lesion segmentation approaches have leveraged artificial intelligence (AI), they often remain proprietary and difficult to adopt. As an open-source tool, we present LST-AI, an advanced deep learning-based extension of LST that consists of an ensemble of three 3D U-Nets. LST-AI explicitly addresses the imbalance between white matter (WM) lesions and non-lesioned WM. It employs a composite loss function incorporating binary cross-entropy and Tversky loss to improve segmentation of the highly heterogeneous MS lesions. We train the network ensemble on 491 MS pairs of T1-weighted and FLAIR images, collected in-house from a 3T MRI scanner, and expert neuroradiologists manually segmented the utilized lesion maps for training. LST-AI also includes a lesion location annotation tool, labeling lesions as periventricular, infratentorial, and juxtacortical according to the 2017 McDonald criteria, and, additionally, as subcortical. We conduct evaluations on 103 test cases consisting of publicly available data using the Anima segmentation validation tools and compare LST-AI with several publicly available lesion segmentation models. Our empirical analysis shows that LST-AI achieves superior performance compared to existing methods. Its Dice and F1 scores exceeded 0.62, outperforming LST, SAMSEG (Sequence Adaptive Multimodal SEGmentation), and the popular nnUNet framework, which all scored below 0.56. Notably, LST-AI demonstrated exceptional performance on the MSSEG-1 challenge dataset, an international WM lesion segmentation challenge, with a Dice score of 0.65 and an F1 score of 0.63-surpassing all other competing models at the time of the challenge. With increasing lesion volume, the lesion detection rate rapidly increased with a detection rate of >75% for lesions with a volume between 10 mm3 and 100 mm3. Given its higher segmentation performance, we recommend that research groups currently using LST transition to LST-AI. To facilitate broad adoption, we are releasing LST-AI as an open-source model, available as a command-line tool, dockerized container, or Python script, enabling diverse applications across multiple platforms.

Effects of sleep fragmentation on white matter pathology in a rat model of cerebral small vessel disease

STUDY OBJECTIVES

Mounting evidence indicated the correlation between sleep and cerebral small vessel disease (CSVD). However, little is known about the exact causality between poor sleep and white matter injury, a typical signature of CSVD, as well as the underlying mechanisms.

METHODS

Spontaneously hypertensive rats (SHR) and control Wistar Kyoto rats were subjected to sleep fragmentation (SF) for 16 weeks. The effects of chronic sleep disruption on the deep white matter and cognitive performance were observed.

RESULTS

SHR were validated as a rat model for CSVD. Fragmented sleep induced strain-dependent white matter abnormalities, characterized by reduced myelin integrity, impaired oligodendrocytes precursor cells (OPC) maturation and pro-inflammatory microglial polarization. Partially reversible phenotypes of OPC and microglia were observed in parallel following sleep recovery.

CONCLUSIONS

Long-term SF-induced pathological effects on the deep white matter in a rat model of CSVD. The pro-inflammatory microglial activation and the block of OPC maturation may be involved in the mechanisms linking sleep to white matter injury.

LST-AI: a Deep Learning Ensemble for Accurate MS Lesion Segmentation

Automated segmentation of brain white matter lesions is crucial for both clinical assessment and scientific research in multiple sclerosis (MS). Over a decade ago, we introduced an engineered lesion segmentation tool, LST. While recent lesion segmentation approaches have leveraged artificial intelligence (AI), they often remain proprietary and difficult to adopt. As an open-source tool, we present LST-AI, an advanced deep learning-based extension of LST that consists of an ensemble of three 3D-UNets. LST-AI explicitly addresses the imbalance between white matter (WM) lesions and non-lesioned WM. It employs a composite loss function incorporating binary cross-entropy and Tversky loss to improve segmentation of the highly heterogeneous MS lesions. We train the network ensemble on 491 MS pairs of T1w and FLAIR images, collected in-house from a 3T MRI scanner, and expert neuroradiologists manually segmented the utilized lesion maps for training. LST-AI additionally includes a lesion location annotation tool, labeling lesion location according to the 2017 McDonald criteria (periventricular, infratentorial, juxtacortical, subcortical). We conduct evaluations on 103 test cases consisting of publicly available data using the Anima segmentation validation tools and compare LST-AI with several publicly available lesion segmentation models. Our empirical analysis shows that LST-AI achieves superior performance compared to existing methods. Its Dice and F1 scores exceeded 0.62, outperforming LST, SAMSEG (Sequence Adaptive Multimodal SEGmentation), and the popular nnUNet framework, which all scored below 0.56. Notably, LST-AI demonstrated exceptional performance on the MSSEG-1 challenge dataset, an international WM lesion segmentation challenge, with a Dice score of 0.65 and an F1 score of 0.63-surpassing all other competing models at the time of the challenge. With increasing lesion volume, the lesion detection rate rapidly increased with a detection rate of >75% for lesions with a volume between 10mm3 and 100mm3. Given its higher segmentation performance, we recommend that research groups currently using LST transition to LST-AI. To facilitate broad adoption, we are releasing LST-AI as an open-source model, available as a command-line tool, dockerized container, or Python script, enabling diverse applications across multiple platforms.

Early detection of white matter hyperintensities using SHIVA-WMH detector

White matter hyperintensities (WMHs) are well-established markers of cerebral small vessel disease, and are associated with an increased risk of stroke, dementia, and mortality. Although their prevalence increases with age, small and punctate WMHs have been reported with surprisingly high frequency even in young, neurologically asymptomatic adults. However, most automated methods to segment WMH published to date are not optimized for detecting small and sparse WMH. Here we present the SHIVA-WMH tool, a deep-learning (DL)-based automatic WMH segmentation tool that has been trained with manual segmentations of WMH in a wide range of WMH severity. We show that it is able to detect WMH with high efficiency in subjects with only small punctate WMH as well as in subjects with large WMHs (i.e., with confluency) in evaluation datasets from three distinct databases: magnetic resonance imaging-Share consisting of young university students, MICCAI 2017 WMH challenge dataset consisting of older patients from memory clinics, and UK Biobank with community-dwelling middle-aged and older adults. Across these three cohorts with a wide-ranging WMH load, our tool achieved voxel-level and individual lesion cluster-level Dice scores of 0.66 and 0.71, respectively, which were higher than for three reference tools tested: the lesion prediction algorithm implemented in the lesion segmentation toolbox (LPA: Schmidt), PGS tool, a DL-based algorithm and the current winner of the MICCAI 2017 WMH challenge (Park et al.), and HyperMapper tool (Mojiri Forooshani et al.), another DL-based method with high reported performance in subjects with mild WMH burden. Our tool is publicly and openly available to the research community to facilitate investigations of WMH across a wide range of severity in other cohorts, and to contribute to our understanding of the emergence and progression of WMH.

Automatic segmentation of white matter hyperintensities in T2-FLAIR with AQUA: A comparative validation study against conventional methods

White matter hyperintensities (WMHs) are lesions in the white matter of the brain that are associated with cognitive decline and an increased risk of dementia. The manual segmentation of WMHs is highly time-consuming and prone to intra- and inter-variability. Therefore, automatic segmentation approaches are gaining attention as a more efficient and objective means to detect and monitor WMHs. In this study, we propose AQUA, a deep learning model designed for fully automatic segmentation of WMHs from T2-FLAIR scans, which improves upon our previous study for small lesion detection and incorporating a multicenter approach. AQUA implements a two-dimensional U-Net architecture and uses patch-based training. Additionally, the network was modified to include Bottleneck Attention Module on each convolutional block of both the encoder and decoder to enhance performance for small-sized WMH. We evaluated the performance and robustness of AQUA by comparing it with five well-known supervised and unsupervised methods for automatic segmentation of WMHs (LGA, LPA, SLS, UBO, and BIANCA). To accomplish this, we tested these six methods on the MICCAI 2017 WMH Segmentation Challenge dataset, which contains MRI images from 170 elderly participants with WMHs of presumed vascular origin, and assessed their robustness across multiple sites and scanner types. The results showed that AQUA achieved superior performance in terms of spatial (Dice = 0.72) and volumetric (logAVD = 0.10) agreement with the manual segmentation compared to the other methods. While the recall and F1-score were moderate at 0.49 and 0.59, respectively, they improved to 0.75 and 0.82 when excluding small lesions (≤ 6 voxels). Remarkably, despite being trained on a different dataset with different ethnic backgrounds, lesion loads, and scanners, AQUA's results were comparable to the top 10 ranked methods of the MICCAI challenge. The findings suggest that AQUA is effective and practical for automatic segmentation of WMHs from T2-FLAIR scans, which could help identify individuals at risk of cognitive decline and dementia and allow for early intervention and management.

Segmenting white matter hyperintensities on isotropic three-dimensional Fluid Attenuated Inversion Recovery magnetic resonance images: Assessing deep learning tools on a Norwegian imaging database

An important step in the analysis of magnetic resonance imaging (MRI) data for neuroimaging is the automated segmentation of white matter hyperintensities (WMHs). Fluid Attenuated Inversion Recovery (FLAIR-weighted) is an MRI contrast that is particularly useful to visualize and quantify WMHs, a hallmark of cerebral small vessel disease and Alzheimer's disease (AD). In order to achieve high spatial resolution in each of the three voxel dimensions, clinical MRI protocols are evolving to a three-dimensional (3D) FLAIR-weighted acquisition. The current study details the deployment of deep learning tools to enable automated WMH segmentation and characterization from 3D FLAIR-weighted images acquired as part of a national AD imaging initiative. Based on data from the ongoing Norwegian Disease Dementia Initiation (DDI) multicenter study, two 3D models-one off-the-shelf from the NVIDIA nnU-Net framework and the other internally developed-were trained, validated, and tested. A third cutting-edge Deep Bayesian network model (HyperMapp3r) was implemented without any de-novo tuning to serve as a comparison architecture. The 2.5D in-house developed and 3D nnU-Net models were trained and validated in-house across five national collection sites among 441 participants from the DDI study, of whom 194 were men and whose average age was (64.91 +/- 9.32) years. Both an external dataset with 29 cases from a global collaborator and a held-out subset of the internal data from the 441 participants were used to test all three models. These test sets were evaluated independently. The ground truth human-in-the-loop segmentation was compared against five established WMH performance metrics. The 3D nnU-Net had the highest performance out of the three tested networks, outperforming both the internally developed 2.5D model and the SOTA Deep Bayesian network with an average dice similarity coefficient score of 0.76 +/- 0.16. Our findings demonstrate that WMH segmentation models can achieve high performance when trained exclusively on FLAIR input volumes that are 3D volumetric acquisitions. Single image input models are desirable for ease of deployment, as reflected in the current embedded clinical research project. The 3D nnU-Net had the highest performance, which suggests a way forward for our need to automate WMH segmentation while also evaluating performance metrics during on-going data collection and model retraining.

Adherence to a healthy lifestyle and brain structural imaging markers

Previous research has linked specific modifiable lifestyle factors to age-related cognitive decline in adults. Little is known about the potential role of an overall healthy lifestyle in brain structure. We examined the association of adherence to a healthy lifestyle with a panel of brain structural markers among 2,413 participants in PolyvasculaR Evaluation for Cognitive Impairment and vaScular Events (PRECISE) study in China and 19,822 participants in UK Biobank (UKB). A healthy lifestyle score (0-5) was constructed based on five modifiable lifestyle factors: diet, physical activity, smoking, alcohol consumption, and body mass index. Validated multimodal neuroimaging markers were derived from brain magnetic resonance imaging. In the cross-sectional analysis of PRECISE, participants who adopted four or five low-risk lifestyle factors had larger total brain volume (TBV; β = 0.12, 95% CI: - 0.02, 0.26; p-trend = 0.05) and gray matter volume (GMV; β = 0.16, 95% CI: 0.01, 0.30; p-trend = 0.05), smaller white matter hyperintensity volume (WMHV; β = - 0.35, 95% CI: - 0.50, - 0.20; p-trend < 0.001) and lower odds of lacune (Odds Ratio [OR] = 0.48, 95% CI: 0.22, 1.08; p-trend = 0.03), compared to those with zero or one low-risk factors. Meanwhile, in the prospective analysis in UKB (with a median of 7.7 years' follow-up), similar associations were observed between the number of low-risk lifestyle factors (4-5 vs. 0-1) and TBV (β = 0.22, 95% CI: 0.16, 0.28; p-trend < 0.001), GMV (β = 0.26, 95% CI: 0.21, 0.32; p-trend < 0.001), white matter volume (WMV; β = 0.08, 95% CI: 0.01, 0.15; p-trend = 0.001), hippocampus volume (β = 0.15, 95% CI: 0.08, 0.22; p-trend < 0.001), and WMHV burden (β = - 0.23, 95% CI: - 0.29, - 0.17; p-trend < 0.001). Those with four or five low-risk lifestyle factors showed approximately 2.0-5.8 years of delay in aging of brain structure. Adherence to a healthier lifestyle was associated with a lower degree of neurodegeneration-related brain structural markers in middle-aged and older adults.

Ensemble learning via supervision augmentation for white matter hyperintensity segmentation

Since the ambiguous boundary of the lesion and inter-observer variability, white matter hyperintensity segmentation annotations are inherently noisy and uncertain. On the other hand, the high capacity of deep neural networks (DNN) enables them to overfit labels with noise and uncertainty, which may lead to biased models with weak generalization ability. This challenge has been addressed by leveraging multiple annotations per image. However, multiple annotations are often not available in a real-world scenario. To mitigate the issue, this paper proposes a supervision augmentation method (SA) and combines it with ensemble learning (SA-EN) to improve the generalization ability of the model. SA can obtain diverse supervision information by estimating the uncertainty of annotation in a real-world scenario that per image have only one ambiguous annotation. Then different base learners in EN are trained with diverse supervision information. The experimental results on two white matter hyperintensity segmentation datasets demonstrate that SA-EN gets the optimal accuracy compared with other state-of-the-art ensemble methods. SA-EN is more effective on small datasets, which is more suitable for medical image segmentation with few annotations. A quantitative study is presented to show the effect of ensemble size and the effectiveness of the ensemble model. Furthermore, SA-EN can capture two types of uncertainty, aleatoric uncertainty modeled in SA and epistemic uncertainty modeled in EN.

Higher-resolution quantification of white matter hypointensities by large-scale transfer learning from 2D images on the JPSC-AD cohort

White matter lesions (WML) commonly occur in older brains and are quantifiable on MRI, often used as a biomarker in Aging research. Although algorithms are regularly proposed that identify these lesions from T2‐fluid‐attenuated inversion recovery (FLAIR) sequences, none so far can estimate lesions directly from T1‐weighted images with acceptable accuracy. Since 3D T1 is a polyvalent and higher‐resolution sequence, it could be beneficial to obtain the distribution of WML directly from it. However a serious difficulty, both for algorithms and human, can be found in the ambiguities of brain signal intensity in T1 images. This manuscript shows that a cross‐domain ConvNet (Convolutional Neural Network) approach can help solve this problem. Still, this is non‐trivial, as it would appear to require a large and varied dataset (for robustness) labelled at the same high resolution (for spatial accuracy). Instead, our model was taught from two‐dimensional FLAIR images with a loss function designed to handle the super‐resolution need. And crucially, we leveraged a very large training set for this task, the recently assembled, multi‐sites Japan Prospective Studies Collaboration for Aging and Dementia (JPSC‐AD) cohort. We describe the two‐step procedure that we followed to handle such a large number of imperfectly labeled samples. A large‐scale accuracy evaluation conducted against FreeSurfer 7, and a further visual expert rating revealed that WML segmentation from our ConvNet was consistently better. Finally, we made a directly usable software program based on that trained ConvNet model, available at https://github.com/bthyreau/deep-T1-WMH.