White matter hyperintensities (WMH) increase with age, with marked interindividual variation. There is a need for normative data by age and sex, to improve individualized WMH burden assessment. In this study, we pooled cross-sectional data from 15 population-based cohorts (14,876 nondemented individuals, age 18-97 years), through the Meta VCI Map consortium. Whole brain and tract-specific MRI-assessed WMH volumes were calculated in MNI-152 space. We used quantile regression to create centile curves of WMH volume versus age, stratified by sex. Total WMH volume and interindividual variance increased exponentially with age for both sexes, with females showing higher WMH volumes. WMH volume increase with aging was not uniform across the white matter, but instead followed one of three different patterns depending on location. Age- and sex-specific normative data for total and regional WMH volumes were created. Our study provides detailed information on the normal distribution of total and regional WMH volumes across adulthood. The normative data enable a quantitative approach to interpreting total and regional WMH volumes in clinical practice and research settings.

Molecular neuroimaging techniques, like PET and SPECT, offer invaluable insights into the brain’s in-vivo biology and its dysfunction in neuropsychiatric patients. However, the transition of molecular neuroimaging into diagnostics and precision medicine has been limited to a few clinical applications, hindered by issues like practical feasibility, high costs, and high between-subject heterogeneity of neuroimaging measures. In this study, we explore the use of normative modelling (NM) to identify individual patient alterations by describing the physiological variability of molecular functions. NM potentially addresses challenges such as small sample sizes and diverse acquisition protocols typical of molecular neuroimaging studies. We applied NM to two PET radiotracers targeting the dopaminergic system ([11C]-(+)-PHNO and [18F]FDOPA) to create a reference-cohort model of healthy controls. The models were subsequently utilized on different independent cohorts of patients with psychosis in different disease stages and treatment outcomes. Our results showed that patients with psychosis exhibited a higher degree of extreme deviations (~3-fold increase) than controls, although this pattern was heterogeneous, with minimal overlap of extreme deviations topology (max 20%). We also confirmed that striatal [18F]FDOPA signal, when referenced to a normative distribution, can predict treatment response (striatal AUC ROC: 0.77–0.83). In conclusion, our results indicate that normative modelling can be effectively applied to molecular neuroimaging after proper harmonization, enabling insights into disease mechanisms and advancing precision medicine. In addition, the method is valuable in understanding the heterogeneity of patient populations and can contribute to maximising cost efficiency in studies aimed at comparing cases and controls.

BACKGROUND:White matter hyperintensities (WMHs) of presumed vascular origin are common in the elderly and are associated with vascular risk factors. There is evidence that vascular risk factors, in particular hypertension, are associated with WMH in particular locations of the white matter. However, it remains unclear whether this is true for all risk factors and whether signature WMH locations differ between risk factors. We aimed to identify WMH locations associated with vascular risk factors in community-dwelling individuals. METHODS:We pooled cross-sectional data from 16 population-based cohorts (15 653 individuals; mean age, 64.2±11.8 years; 52.2% female) through the Meta VCI Map Consortium. We quantified associations between WMH volumes in 50 white matter regions and 6 vascular risk factors using linear mixed models. Analyses were corrected for age, sex, study site, and total WMH volume. RESULTS:Hypertension (B=0.141; P<0.001), smoking (B=0.096; P<0.001), diabetes (B=0.059; P<0.001), and history of vascular disease (B=0.056; P=0.034) were significantly associated with higher total WMH volume, whereas obesity (B=0.023; P=0.139) and hypercholesterolemia (B=0.009; P=0.531) were not. After correcting for total WMH volume, hypertension was associated with WMH volume in 10 regions (ie, bilateral external capsule, superior longitudinal fasciculus, superior corona radiata, anterior limb of the internal capsule, left anterior corona radiata, and left superior fronto-occipital fasciculus), smoking (body corpus callosum), diabetes (genu corpus callosum), and obesity (left inferior fronto-occipital fasciculus), each with one region. CONCLUSIONS:Hypertension has a signature WMH pattern, whereas associations between other vascular risk factors and regional WMH volumes seem to be mainly explained by a global increase in WMH rather than region-specific effects.

Brain white matter disruptions have been implicated in contributing to fatigue, brain fog and other central symptoms commonly reported in inflammatory diseases. In this study, we included 252 RA patients with 756 age and sex matched controls and 240 UC patients with 720 age and sex matched controls using the UK Biobank imaging dataset. We looked for differences in total volume of white matter hyperintensities (WMH) between patients compared to controls. Then, using voxelwise analysis, we explored the spatial distribution of these white matter hyperintensities and differences in these between patients and controls and between disease groups. A significantly higher volume of WMH was observed in both the RA (p = 1.9 × 10-8, β =  - 0.36, 95% CI =  - 0.48, - 0.23) and UC (p = 0.003, β =  - 0.18 95% CI =  - 0.31, - 0.06) patients compared to their respective control groups. Voxelwise analysis revealed only a small cluster of RA associated WMH compared to controls. These results indicate an increased risk of white matter hyperintensities in patients with RA and UC. These findings help quantify the effect of inflammation from autoimmune diseases on cerebrovascular health and white matter integrity.

In this paper, we study pseudo-labelling. Pseudo-labelling employs raw inferences on unlabelled data as pseudo-labels for self-training. We elucidate the empirical successes of pseudo-labelling by establishing a link between this technique and the Expectation Maximisation algorithm. Through this, we realise that the original pseudo-labelling serves as an empirical estimation of its more comprehensive underlying formulation. Following this insight, we present a full generalisation of pseudo-labels under Bayes' theorem, termed Bayesian Pseudo Labels. Subsequently, we introduce a variational approach to generate these Bayesian Pseudo Labels, involving the learning of a threshold to automatically select high-quality pseudo labels. In the remainder of the paper, we showcase the applications of pseudo-labelling and its generalised form, Bayesian Pseudo-Labelling, in the semi-supervised segmentation of medical images. Specifically, we focus on: 1) 3D binary segmentation of lung vessels from CT volumes; 2) 2D multi-class segmentation of brain tumours from MRI volumes; 3) 3D binary segmentation of whole brain tumours from MRI volumes; and 4) 3D binary segmentation of prostate from MRI volumes. We further demonstrate that pseudo-labels can enhance the robustness of the learned representations. The code is released in the following GitHub repository: https://github.com/moucheng2017/EMSSL

BackgroundRheumatoid arthritis (RA) and ulcerative colitis (UC) are two autoimmune diseases where patients report high levels of fatigue, pain, and depression. The effect of systemic inflammation from these diseases is likely affecting the brain, however, it is unknown whether there are measurable neuroanatomical changes and whether these are a contributing factor to these central symptoms.MethodsWe included 258 RA patients with 774 age and sex matched controls and 249 UC patients with 747 age and sex matched controls in a case control study utilizing the UK Biobank dataset. We used imaging derived phenotypes (IDPs) to determine whether there were differences in (1) hippocampal volume and (2) additional subcortical brain volumes between patients compared to controls and if there were common regions affected between these two diseases.ResultsPatients with UC had moderately smaller hippocampi compared to age and sex matched controls (difference: 134.15 mm3, SD ± 64.76, p = 0.035). This result was not seen in RA patients. RA patients had a significantly smaller amygdala volume than age and sex matched controls (difference: 91.27 mm3, SD ± 30.85, p = 0.0021, adjusted p = 0.012). This result was not seen in UC patients. All other subcortical structures analyzed were comparable between the patients and control groups.ConclusionThese results indicate there are subcortical brain differences between UC, RA and controls but different regions of the limbic system are preferentially affected by UC and RA. This study may provide evidence for different neurodegenerative mechanisms in distinct autoimmune diseases.

Imaging markers of cerebral small vessel disease provide valuable information on brain health, but their manual assessment is time-consuming and hampered by substantial intra- and interrater variability. Automated rating may benefit biomedical research, as well as clinical assessment, but diagnostic reliability of existing algorithms is unknown. Here, we present the results of the VAscular Lesions DetectiOn and Segmentation (Where is VALDO?) challenge that was run as a satellite event at the international conference on Medical Image Computing and Computer Aided Intervention (MICCAI) 2021. This challenge aimed to promote the development of methods for automated detection and segmentation of small and sparse imaging markers of cerebral small vessel disease, namely enlarged perivascular spaces (EPVS) (Task 1), cerebral microbleeds (Task 2) and lacunes of presumed vascular origin (Task 3) while leveraging weak and noisy labels. Overall, 12 teams participated in the challenge proposing solutions for one or more tasks (4 for Task 1 - EPVS, 9 for Task 2 - Microbleeds and 6 for Task 3 - Lacunes). Multi-cohort data was used in both training and evaluation. Results showed a large variability in performance both across teams and across tasks, with promising results notably for Task 1 - EPVS and Task 2 - Microbleeds and not practically useful results yet for Task 3 - Lacunes. It also highlighted the performance inconsistency across cases that may deter use at an individual level, while still proving useful at a population level.

In this paper, we study pseudo-labelling. Pseudo-labelling employs raw inferences on unlabelled data as pseudo-labels for self-training. We elucidate the empirical successes of pseudo-labelling by establishing a link between this technique and the Expectation Maximisation algorithm. Through this, we realise that the original pseudo-labelling serves as an empirical estimation of its more comprehensive underlying formulation. Following this insight, we present a full generalisation of pseudo-labels under Bayes' theorem, termed Bayesian Pseudo Labels. Subsequently, we introduce a variational approach to generate these Bayesian Pseudo Labels, involving the learning of a threshold to automatically select high-quality pseudo labels. In the remainder of the paper, we showcase the applications of pseudo-labelling and its generalised form, Bayesian Pseudo-Labelling, in the semi-supervised segmentation of medical images. Specifically, we focus on: 1) 3D binary segmentation of lung vessels from CT volumes; 2) 2D multi-class segmentation of brain tumours from MRI volumes; 3) 3D binary segmentation of whole brain tumours from MRI volumes; and 4) 3D binary segmentation of prostate from MRI volumes. We further demonstrate that pseudo-labels can enhance the robustness of the learned representations. The code is released in the following GitHub repository: https://github.com/moucheng2017/EMSSL

Semi-supervised learning (SSL) is a promising machine learning paradigm to address the ubiquitous issue of label scarcity in medical imaging. The state-of-the-art SSL methods in image classification utilise consistency regularisation to learn unlabelled predictions which are invariant to input level perturbations. However, image level perturbations violate the cluster assumption in the setting of segmentation. Moreover, existing image level perturbations are hand-crafted which could be sub-optimal. In this paper, we propose MisMatch, a semi-supervised segmentation framework based on the consistency between paired predictions which are derived from two differently learnt morphological feature perturbations. MisMatch consists of an encoder and two decoders. One decoder learns positive attention for foreground on unlabelled data thereby generating dilated features of foreground. The other decoder learns negative attention for foreground on the same unlabelled data thereby generating eroded features of foreground. We normalise the paired predictions of the decoders, along the batch dimension. A consistency regularisation is then applied between the normalised paired predictions of the decoders. We evaluate MisMatch on four different tasks. Firstly, we develop a 2D U-net based MisMatch framework and perform extensive cross-validation on a CT-based pulmonary vessel segmentation task and show that MisMatch statistically outperforms state-of-the-art semi-supervised methods. Secondly, we show that 2D MisMatch outperforms state-of-the-art methods on an MRI-based brain tumour segmentation task. We then further confirm that 3D V-net based MisMatch outperforms its 3D counterpart based on consistency regularisation with input level perturbations, on two different tasks including, left atrium segmentation from 3D CT images and whole brain tumour segmentation from 3D MRI images. Lastly, we find that the performance improvement of MisMatch over the baseline might originate from its better calibration. This also implies that our proposed AI system makes safer decisions than the previous methods.

Molecular neuroimaging techniques, like PET and SPECT, offer invaluable insights into the brain's in-vivo biology and its dysfunction in neuropsychiatric patients. However, the transition of molecular neuroimaging into diagnostics and precision medicine has been limited to a few clinical applications, hindered by issues like practical feasibility and high costs. In this study, we explore the use of normative modelling (NM) for molecular neuroimaging to identify individual patient deviations from a reference cohort of subjects. NM potentially addresses challenges such as small sample sizes and diverse acquisition protocols that are typical of molecular neuroimaging studies. We applied NM to two PET radiotracers targeting the dopaminergic system ([11C]-(+)-PHNO and [18F]FDOPA) to create a normative model to reference groups of controls. The models were subsequently utilized on various independent cohorts of patients experiencing psychosis. These cohorts were characterized by differing disease stages, treatment responses, and the presence or absence of matched controls. Our results showed that patients exhibited a higher degree of extreme deviations (~3-fold increase) than controls, although this pattern was heterogeneous, with minimal overlap in extreme deviations topology (max 20%). We also confirmed the value of striatal [18F]FDOPA signal to predict treatment response (striatal AUC ROC: 0.77-0.83). Methodologically, we highlighted the importance of data harmonization before data aggregation. In conclusion, normative modelling can be effectively applied to molecular neuroimaging after proper harmonization, enabling insights into disease mechanisms and advancing precision medicine. The method is valuable in understanding the heterogeneity of patient populations and can contribute to maximising cost efficiency in studies aimed at comparing cases and controls.### Competing Interest StatementM.d.G. is an employee of GSK, GSK had no role in the design of this study. R.A.M. has received speaker/consultancy fees from Karuna, Janssen, Boehringer Ingelheim, and Otsuka, and co-directs a company that designs digital resources to support treatment of mental illness. F.B. has received consulting fees from Petalouda Therapeutics and has been an employee at Compass Pathways. AE has received consulting fees from Leal Therapeutics.### Funding StatementThis study represents independent research funded by the NIHR Maudsley Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King's College London. The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care. A.G. is supported by the KCL-funded CDT in Data-Driven Health; this represents independent research partly funded by the NIHR Maudsley's Biomedical Research Centre (BRC) at the South London and Maudsley NHS Foundation Trust and partly funded by GSK. D.M., G.N., R.E., O.D., F.T., and S.C.R.W. are supported by the NIHR Maudsley's Biomedical Research Centre at the South London and Maudsley NHS Trust. MV is supported by the Italian National Centre for HPC, BIG DATA AND QUANTUM COMPUTING (Project no. CN00000013 CN1), the PNR Italian National Grant DIGITAL LIFELONG PREVENTION (Project no PNC0000002_DARE), and by Wellcome Trust Digital Award (no. 215747/Z/19/Z). M.d.G. is an employee of GSK, GSK had no role in the design of this study.### Author DeclarationsI confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.YesThe details of the IRB/oversight body that provided approval or exemption for the research described are given below:All the research protocols for data acquisitions were approved by local ethics committees and institutional revision boards including the Institute of Psychiatry, King's College, London, England, research ethics committee; the South London and Maudsley/Institute of Psychiatry NHS Trust, London-West London & GTAC Research Ethics Committee; the Administration of Radioactive Substances Advisory Committee (ARSAC); the Hammersmith Research Ethics Committee; the East of England-Cambridge East NHS Research Ethics Committee; Seoul National University Hospital, Seoul, Korea.I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.YesI understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).YesI have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.YesThe data that support the findings of this study are available from The NeurOimaging DatabasE (NODE) repository (https://maudsleybrc.nihr.ac.uk/research/precision-psychiatr y/neuroimaging/neuroimaging-database-node/) but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission by the data controller institutions, by contacting the support team (node.information@kcl.ac.uk) or the author Dr. Giovanna Nordio (giovanna.nordio@kcl.ac.uk).

A comprehensive search of published literature in brain volumetry was conducted in three autoimmune diseases - systemic lupus erythematosus (SLE), rheumatoid arthritis (RA), and ulcerative colitis (UC) - with the intention of performing a meta-analysis of published data. Due to the lack of data in RA and UC, the reported meta-analysis was limited to SLE. The MEDLINE database was searched for studies from 1988 to March 2022. A total of 175 papers met the initial inclusion criteria, and 16 were included in a random-effects meta-analysis. The reduction in the number of papers included in the final analysis is primarily due to the lack of overlap in measured and reported brain regions. A significantly lower volume was seen in patients with SLE in the hippocampus, corpus callosum, and total gray matter volume measurements as compared to age- and sex-matched controls. There were not enough studies to perform a meta-analysis for RA and UC; instead, we include a summary of published volumetric studies. The meta-analyses revealed structural brain abnormalities in patients with SLE, suggesting that lower global brain volumes are associated with disease status. This volumetric difference was seen in both the hippocampus and corpus callosum and total gray matter volume measurements. These results indicate both gray and white matter involvements in SLE and suggest there may be both localized and global reductions in brain volume.

This paper concerns pseudo labelling in segmentation. Our contribution is fourfold. Firstly, we present a new formulation of pseudo-labelling as an Expectation-Maximization (EM) algorithm for clear statistical interpretation. Secondly, we propose a semi-supervised medical image segmentation method purely based on the original pseudo labelling, namely SegPL. We demonstrate SegPL is a competitive approach against state-of-the-art consistency regularisation based methods on semi-supervised segmentation on a 2D multi-class MRI brain tumour segmentation task and a 3D binary CT lung vessel segmentation task. The simplicity of SegPL allows less computational cost comparing to prior methods. Thirdly, we demonstrate that the effectiveness of SegPL may originate from its robustness against out-of-distribution noises and adversarial attacks. Lastly, under the EM framework, we introduce a probabilistic generalisation of SegPL via variational inference, which learns a dynamic threshold for pseudo labelling during the training. We show that SegPL with variational inference can perform uncertainty estimation on par with the gold-standard method Deep Ensemble.

We propose MisMatch, a novel consistency-driven semi-supervised segmentation framework which produces predictions that are invariant to learnt feature perturbations. MisMatch consists of an encoder and a two-head decoders. One decoder learns positive attention to the foreground regions of interest (RoI) on unlabelled images thereby generating dilated features. The other decoder learns negative attention to the foreground on the same unlabelled images thereby generating eroded features. We then apply a consistency regularisation on the paired predictions. MisMatch outperforms state-of-the-art semi-supervised methods on a CT-based pulmonary vessel segmentation task and a MRI-based brain tumour segmentation task. In addition, we show that the effectiveness of MisMatch comes from better model calibration than its supervised learning counterpart.

OBJECTIVES:To assess non-invasive imaging for detection and quantification of gland structure, inflammation and function in patients with primary Sjogren's syndrome (pSS) using PET-CT with 11C-Methionine (11C-MET; radiolabelled amino acid), and 18F-fluorodeoxyglucose (18F-FDG; glucose uptake marker), to assess protein synthesis and inflammation, respectively; multiparametric MRI evaluated salivary gland structural and physiological changes. METHODS:In this imaging/clinical/histology comparative study (GSK study 203818; NCT02899377) patients with pSS and age- and sex-matched healthy volunteers underwent MRI of the salivary glands and 11C-MET PET-CT. Patients also underwent 18F-FDG PET-CT and labial salivary gland biopsies. Clinical and biomarker assessments were performed. Primary endpoints were semi-quantitative parameters of 11C-MET and 18F-FDG uptake in submandibular and parotid salivary glands and quantitative MRI measures of structure and inflammation. Clinical and minor salivary gland histological parameter correlations were explored. RESULTS:Twelve patients with pSS and 13 healthy volunteers were included. Lower 11C-MET uptake in parotid, submandibular and lacrimal glands, lower submandibular gland volume, higher MRI fat fraction, and lower pure diffusion in parotid and submandibular glands were observed in patients vs healthy volunteer, consistent with reduced synthetic function. Disease duration correlated positively with fat fraction and negatively with 11C-MET and 18F-FDG uptake, consistent with impaired function, inflammation and fatty replacement over time. Lacrimal gland 11C-MET uptake positively correlated with tear flow in patients, and parotid gland 18F-FDG uptake positively correlated with salivary gland CD20+ B-cell infiltration. CONCLUSION:Molecular imaging and MRI may be useful tools to non-invasively assess loss of glandular function, increased glandular inflammation and fat accumulation in pSS.

The lack of labels is one of the fundamental constraints in deep learning based methods for image classification and segmentation, especially in applications such as medical imaging. Semi-supervised learning (SSL) is a promising method to address the challenge of label scarcity. The state-of-the-art SSL methods utilise consistency regularisation to learn unlabelled predictions which are invariant to perturbations on the prediction confidence. However, such SSL approaches rely on hand-crafted augmentation techniques which could be sub-optimal. In this paper, we propose MisMatch, a novel consistency based semi-supervised segmentation method. MisMatch automatically learns to produce paired predictions with increased and decreased confidences. MisMatch consists of an encoder and two decoders. One decoder learns positive attention for regions of interest (RoI) on unlabelled data thereby generating higher confidence predictions of RoI. The other decoder learns negative attention for RoI on the same unlabelled data thereby generating lower confidence predictions. We then apply a consistency regularisation between the paired predictions of the decoders. For evaluation, we first perform extensive cross-validation on a CTbased pulmonary vessel segmentation task and show that MisMatch statistically outperforms state-of-the-art semisupervised methods when only 6.25% of the total labels are used. Furthermore MisMatch performance using 6.25% of the total labels is comparable to state-of-the-art methods that utilise all available labels. In a second experiment, MisMatch outperforms state-of-the-art methods on an MRIbased brain tumour segmentation task.

This work presents a single-step deep-learning framework for longitudinal image analysis, coined Segis-Net. To optimally exploit information available in longitudinal data, this method concurrently learns a multi-class segmentation and nonlinear registration. Segmentation and registration are modeled using a convolutional neural network and optimized simultaneously for their mutual benefit. An objective function that optimizes spatial correspondence for the segmented structures across time-points is proposed. We applied Segis-Net to the analysis of white matter tracts from N=8045 longitudinal brain MRI datasets of 3249 elderly individuals. Segis-Net approach showed a significant increase in registration accuracy, spatio-temporal segmentation consistency, and reproducibility compared with two multistage pipelines. This also led to a significant reduction in the sample-size that would be required to achieve the same statistical power in analyzing tract-specific measures. Thus, we expect that Segis-Net can serve as a new reliable tool to support longitudinal imaging studies to investigate macro- and microstructural brain changes over time.

VALDO Challenge 2021 Training dataset This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-sa/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA. This dataset corresponds to the training data of the VALDO Challenge 2021 - For more information about the challenge and to participate, please see website In order to register a team, remember to fill in the registration form The challenge is separated in 3 tasks Task 1 - Perivascular spaces segmentation Task 2 - Cerebral micro bleeds segmentation Task 3 - Lacunes segmentation Details of the contents and organisation of the data is described in the README.md file Data reference and acknowledgement When using this data, we kindly ask for the following funding sources to be acknowledged: Wellcome Trust (082464/Z/07/Z), British Heart Foundation (SP/07/001/23603, PG/08/103, PG/12/29/29497 and CS/13/1/30327), Erasmus MC University Medical Center, the Erasmus University Rotterdam, the Netherlands Organization for Scientific Research (NWO) Grant 918-46-615, the Netherlands Organization for Health Research and Development (ZonMW), the Research Institute for Disease in the Elderly (RIDE), and the European Union Seventh Framework Programme (FP7/2007–2013) under grant agreement No. 601055, VPHDARE@IT, the Dutch Technology Foundation STW For all publications, please add references noted below in the reference section When using data from Task 2, please add "for the ALFA Study" as corporate author in your publication and indicate the following list of contributors: Müge Akinci, Annabella Beteta, Raffaele Cacciaglia, Alba Cañas, Irene Cumplido, Carme Deulofeu, Ruth Dominguez, Maria Emilio, Carles Falcón, Karine Fauria, Sherezade Fuentes, Juan Domingo Gispert, Oriol Grau-Rivera, José M. González-de-Echávarri, Laura Hernandez, Gema Huesa, Jordi Huguet, Iva Knezevic, Eider M. Arenaza-Urquijo, Eva M Palacios, Paula Marne, Tania Menchón, Marta Milà-Alomà, Carolina Minguillon, José Luis Molinuevo, Grégory Operto, Albina Polo, Gemma Salvadó, Sandra Pradas, Blanca Rodríguez, Aleix Sala-Vila, Gonzalo Sánchez-Benavides, Mahnaz Shekari, Anna Soteras, Marc Suárez-Calvet, Laura Stankeviciute, Marc Vilanova and Natalia Vilor-Tejedor.

To measure the volume of specific image structures, a typical approach is to first segment those structures using a neural network trained on voxel-wise (strong) labels and subsequently compute the volume from the segmentation. A more straightforward approach would be to predict the volume directly using a neural network based regression approach, trained on image-level (weak) labels indicating volume. In this article, we compared networks optimized with weak and strong labels, and study their ability to generalize to other datasets. We experimented with white matter hyperintensity (WMH) volume prediction in brain MRI scans. Neural networks were trained on a large local dataset and their performance was evaluated on four independent public datasets. We showed that networks optimized using only weak labels reflecting WMH volume generalized better for WMH volume prediction than networks optimized with voxel-wise segmentations of WMH. The attention maps of networks trained with weak labels did not seem to delineate WMHs, but highlighted instead areas with smooth contours around or near WMHs. By correcting for possible confounders we showed that networks trained on weak labels may have learnt other meaningful features that are more suited to generalization to unseen data. Our results suggest that for imaging biomarkers that can be derived from segmentations, training networks to predict the biomarker directly may provide more robust results than solving an intermediate segmentation step.

INTRODUCTION:As hearing loss has been identified as an important risk factor for dementia, we aimed to assess the association between hearing loss and microstructural integrity of the brain.METHODS:A total of 1086 dementia-free participants (mean age = 75.2 [standard deviation: 4.9], 61.4% female) of the population-based Atherosclerosis Risk in Communities (ARIC) study underwent hearing assessment (2016-2017) and magnetic resonance imaging of the brain (2011-2013). Microstructural integrity was determined with diffusion tensor imaging. Multivariable linear regression was used to investigate associations between hearing loss and microstructural integrity of different brain regions and white matter (WM) tracts.RESULTS:Hearing loss was associated with lower WM microstructural integrity in the temporal lobe, lower gray matter integrity of the hippocampus, and with lower WM microstructural integrity of the limbic tracts and the uncinate fasciculus.CONCLUSION:Our results demonstrate that hearing loss is indepedently associated with lower microstructural integrity in brain regions that are important for different cognitive processes.