Performance of an Open-Source Large Language Model in Extracting Information from Free-Text Radiology Reports.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To assess the performance of a local open-source large language model (LLM) on various information extraction tasks from real-life emergency brain MRI reports. Materials and Methods All consecutive emergency brain MRI reports written in 2022 from a French quaternary center were retrospectively reviewed. Two radiologists identified MRIs that were performed for headaches. Four radiologists scored reports' conclusions as normal or abnormal. Abnormalities were labeled as either headache-causing or incidental. Vicuna, an open-source LLM, performed the same tasks. Vicuna's performance metrics were evaluated using the radiologists' consensus as the reference standard. Results Among the 2398 reports during the study period, radiologists identified 595 that included headaches in their indication (median age of patients, 35 years [IQR, 26-51], 68% (403/595) female). A positive finding was reported in 227/595 (38%) cases, 136 of which could explain the headache. The LLM had a sensitivity/specificity (95%CI), respectively, of 98% (583/595)(97-99)/99% (1791/1803)(99-100) for detecting the presence of headache in the clinical context, 99% (514/517)(98-100)/99% (68/69)(92-100) for the use of contrast medium injection, 97% (219/227)(93-99)/99% (364/368)(97-100) for study categorization as normal or abnormal and 88% (120/136)(82- 93)/73% (66/91)(62-81) for causal inference between MRI findings and headache. Conclusion An open-source LLM was able to extract information from free-text radiology reports with excellent accuracy without requiring further training. ©RSNA, 2024.

Clinical and neuroimaging predictors of benzodiazepine response in catatonia: A machine learning approach.

Catatonia is a well characterized psychomotor syndrome combining motor, behavioural and neurovegetative signs. Benzodiazepines are the first-choice treatment, effective in 70 % of cases. Currently, the factors associated with benzodiazepine resistance remain unknown. We aimed to develop machine learning models using clinical and neuroimaging data to predict benzodiazepine response in catatonic patients. This study examined a cohort of catatonic patients who underwent standardized clinical evaluation, 3 T brain MRI, and benzodiazepine trial. Based on clinical response, patients were classified as benzodiazepine responders or non-responders. Cortical thickness and regional brain volumes were measured. Two machine learning models (linear model and gradient boosting tree model) were developed to identify predictors of treatment response using clinical, demographic, and neuroimaging data. The cohort included 65 catatonic patients, comprising 30 benzodiazepine responders and 35 non-responders. Using clinical data alone, the linear model achieved 63% precision, 51% recall, a specificity of 61%, and 58% AUC, while the gradient boosting tree (GBT) model attained 46% precision, 60% recall, a specificity of 62% and 64% AUC. Incorporating neuroimaging data improved model performance, with the linear model achieving 66% precision, 57% recall, a specificity of 67%, and 70% AUC, and the GBT model attaining 50% precision, 50% recall, a specificity of 62% and 70% AUC. The integration of imaging data with demographic and clinical information significantly enhanced the predictive performance of the models. The duration of the catatonic syndrome, along with the presence of mitgehen (passive obedience) and immobility/stupor, and the volume of the right medial orbito-frontal cortex emerged as important factors in predicting non-response to benzodiazepines.

The Use of Image Guided Programming to Improve Deep Brain Stimulation Workflows with Directional Leads in Parkinson's Disease.

BACKGROUND: Deep brain stimulation (DBS) of the subthalamic nucleus (STN) is a preferred treatment for parkinsonian patients with severe motor fluctuations. Proper targeting of the STN sensorimotor segment appears to be a crucial factor for success of the procedure. The recent introduction of directional leads theoretically increases stimulation specificity in this challenging area but also requires more precise stimulation parameters. OBJECTIVE: We investigated whether commercially available software for image guided programming (IGP) could maximize the benefits of DBS by informing the clinical standard care (CSC) and improving programming workflows. METHODS: We prospectively analyzed 32 consecutive parkinsonian patients implanted with bilateral directional leads in the STN. Double blind stimulation parameters determined by CSC and IGP were assessed and compared at three months post-surgery. IGP was used to adjust stimulation parameters if further clinical refinement was required. Overall clinical efficacy was evaluated one-year post-surgery. RESULTS: We observed 78% concordance between the two electrode levels selected by the blinded IGP prediction and CSC assessments. In 64% of cases requiring refinement, IGP improved clinical efficacy or reduced mild side effects, predominantly by facilitating the use of directional stimulation (93% of refinements). CONCLUSIONS: The use of image guided programming saves time and assists clinical refinement, which may be beneficial to the clinical standard care for STN-DBS and further improve the outcomes of DBS for PD patients.

A three-dimensional deep learning model for inter-site harmonization of structural MR images of the brain: Extensive validation with a multicenter dataset.

In multicenter MRI studies, pooling the imaging data can introduce site-related variabilities and can therefore bias the subsequent analyses. To harmonize the intensity distributions of brain MR images in a multicenter dataset, unsupervised deep learning methods can be employed. Here, we developed a model based on cycle-consistent adversarial networks for the harmonization of T1-weighted brain MR images. In contrast to previous works, it was designed to process three-dimensional whole-brain images in a stable manner while optimizing computation resources. Using six different MRI datasets for healthy adults (n=1525 in total) with different acquisition parameters, we tested the model in (i) three pairwise harmonizations with site effects of various sizes, (ii) an overall harmonization of the six datasets with different age distributions, and (iii) a traveling-subject dataset. Our results for intensity distributions, brain volumes, image quality metrics and radiomic features indicated that the MRI characteristics at the various sites had been effectively homogenized. Next, brain age prediction experiments and the observed correlation between the gray-matter volume and age showed that thanks to an appropriate training strategy and despite biological differences between the dataset populations, the model reinforced biological patterns. Furthermore, radiologic analyses of the harmonized images attested to the conservation of the radiologic information in the original images. The robustness of the harmonization model (as judged with various datasets and metrics) demonstrates its potential for application in retrospective multicenter studies.

MRI BrainAGE demonstrates increased brain aging in systemic lupus erythematosus patients.

Introduction: Systemic lupus erythematosus (SLE) is an autoimmune connective tissue disease affecting multiple organs in the human body, including the central nervous system. Recently, an artificial intelligence method called BrainAGE (Brain Age Gap Estimation), defined as predicted age minus chronological age, has been developed to measure the deviation of brain aging from a healthy population using MRI. Our aim was to evaluate brain aging in SLE patients using a deep-learning BrainAGE model. Methods: Seventy female patients with a clinical diagnosis of SLE and 24 healthy age-matched control females, were included in this post-hoc analysis of prospectively acquired data. All subjects had previously undergone a 3 T MRI acquisition, a neuropsychological evaluation and a measurement of neurofilament light protein in plasma (NfL). A BrainAGE model with a 3D convolutional neural network architecture, pre-trained on the 3D-T1 images of 1,295 healthy female subjects to predict their chronological age, was applied on the images of SLE patients and controls in order to compute the BrainAGE. SLE patients were divided into 2 groups according to the BrainAGE distribution (high vs. low BrainAGE). Results: BrainAGE z-score was significantly higher in SLE patients than in controls (+0.6 [±1.1] vs. 0 [±1.0], p = 0.02). In SLE patients, high BrainAGE was associated with longer reaction times (p = 0.02), lower psychomotor speed (p = 0.001) and cognitive flexibility (p = 0.04), as well as with higher NfL after adjusting for age (p = 0.001). Conclusion: Using a deep-learning BrainAGE model, we provide evidence of increased brain aging in SLE patients, which reflected neuronal damage and cognitive impairment.

Targeted magnetic resonance imaging (tMRI) of small changes in the T1 and spatial properties of normal or near normal appearing white and gray matter in disease of the brain using divided subtracted inversion recovery (dSIR) and divided reverse subtracted inversion recovery (drSIR) sequences.

This review describes targeted magnetic resonance imaging (tMRI) of small changes in the T1 and the spatial properties of normal or near normal appearing white or gray matter in disease of the brain. It employs divided subtracted inversion recovery (dSIR) and divided reverse subtracted inversion recovery (drSIR) sequences to increase the contrast produced by small changes in T1 by up to 15 times compared to conventional T1-weighted inversion recovery (IR) sequences such as magnetization prepared-rapid acquisition gradient echo (MP-RAGE). This increase in contrast can be used to reveal disease with only small changes in T1 in normal appearing white or gray matter that is not apparent on conventional MP-RAGE, T2-weighted spin echo (T2-wSE) and/or fluid attenuated inversion recovery (T2-FLAIR) images. The small changes in T1 or T2 in disease are insufficient to produce useful contrast with conventional sequences. To produce high contrast dSIR and drSIR sequences typically need to be targeted for the nulling TI of normal white or gray matter, as well as for the sign and size of the change in T1 in these tissues in disease. The dSIR sequence also shows high signal boundaries between white and gray matter. dSIR and drSIR are essentially T1 maps. There is a nearly linear relationship between signal and T1 in the middle domain (mD) of the two sequences which includes T1s between the nulling T1s of the two acquired IR sequences. The drSIR sequence is also very sensitive to reductions in T1 produced by Gadolinium based contrast agents (GBCAs), and when used with rigid body registration to align three-dimensional (3D) isotropic pre and post GBCA images may be of considerable value in showing subtle GBCA enhancement. In serial MRI studies performed at different times, the high signal boundaries generated by dSIR and drSIR sequences can be used with rigid body registration of 3D isotropic images to demonstrate contrast arising from small changes in T1 (without or with GBCA enhancement) as well as small changes in the spatial properties of normal tissues and lesions, such as their site, shape, size and surface. Applications of the sequences in cases of multiple sclerosis (MS) and methamphetamine dependency are illustrated. Using targeted narrow mD dSIR sequences, widespread abnormalities were seen in areas of normal appearing white matter shown with conventional T2-wSE and T2-FLAIR sequences. Understanding of the features of dSIR and drSIR images is facilitated by the use of their T1-bipolar filters; to explain their targeting, signal, contrast, boundaries, T1 mapping and GBCA enhancement. Targeted MRI (tMRI) using dSIR and drSIR sequences may substantially improve clinical MRI of the brain by providing unequivocal demonstration of abnormalities that are not seen with conventional sequences.

Advanced imaging in first episode psychosis: a systematic review.

First-episode psychosis (FEP) is defined as the first occurrence of delusions, hallucinations, or psychic disorganization of significant magnitude, lasting more than 7 days. Evolution is difficult to predict since the first episode remains isolated in one third of cases, while recurrence occurs in another third, and the last third progresses to a schizo-affective disorder. It has been suggested that the longer psychosis goes unnoticed and untreated, the more severe the probability of relapse and recovery. MRI has become the gold standard for imaging psychiatric disorders, especially first episode psychosis. Besides ruling out some neurological conditions that may have psychiatric manifestations, advanced imaging techniques allow for identifying imaging biomarkers of psychiatric disorders. We performed a systematic review of the literature to determine how advanced imaging in FEP may have high diagnostic specificity and predictive value regarding the evolution of disease.

Hypothalamic neuroglial plasticity is regulated by anti-Müllerian hormone and disrupted in polycystic ovary syndrome.

BACKGROUND: Polycystic ovary syndrome (PCOS) is the most common reproductive-endocrine disorder affecting between 5 and 18% of women worldwide. An elevated frequency of pulsatile luteinizing hormone (LH) secretion and higher serum levels of anti-Müllerian hormone (AMH) are frequently observed in women with PCOS. The origin of these abnormalities is, however, not well understood. METHODS: We studied brain structure and function in women with and without PCOS using proton magnetic resonance spectroscopy (MRS) and diffusion tensor imaging combined with fiber tractography. Then, using a mouse model of PCOS, we investigated by electron microscopy whether AMH played a role on the regulation of hypothalamic structural plasticity. FINDINGS: Increased AMH serum levels are associated with increased hypothalamic activity/axonal-glial signalling in PCOS patients. Furthermore, we demonstrate that AMH promotes profound micro-structural changes in the murine hypothalamic median eminence (ME), creating a permissive environment for GnRH secretion. These include the retraction of the processes of specialized AMH-sensitive ependymo-glial cells called tanycytes, allowing more GnRH neuron terminals to approach ME blood capillaries both during the run-up to ovulation and in a mouse model of PCOS. INTERPRETATION: We uncovered a central function for AMH in the regulation of fertility by remodeling GnRH terminals and their tanycytic sheaths, and provided insights into the pivotal role of the brain in the establishment and maintenance of neuroendocrine dysfunction in PCOS. FUNDING: INSERM (U1172), European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement n° 725149), CHU de Lille, France (Bonus H).

European recommendations on practices in pediatric neuroradiology: consensus document from the European Society of Neuroradiology (ESNR), European Society of Paediatric Radiology (ESPR) and European Union of Medical Specialists Division of Neuroradiology (UEMS).

Pediatric neuroradiology is a subspecialty within radiology, with possible pathways to train within the discipline from neuroradiology or pediatric radiology. Formalized pediatric neuroradiology training programs are not available in most European countries. We aimed to construct a European consensus document providing recommendations for the safe practice of pediatric neuroradiology. We particularly emphasize imaging techniques that should be available, optimal site conditions and facilities, recommended team requirements and specific indications and protocol modifications for each imaging modality employed for pediatric neuroradiology studies. The present document serves as guidance to the optimal setup and organization for carrying out pediatric neuroradiology diagnostic and interventional procedures. Clinical activities should always be carried out in full agreement with national provisions and regulations. Continued education of all parties involved is a requisite for preserving pediatric neuroradiology practice at a high level.

Quantitative susceptibility mapping demonstrates different patterns of iron overload in subtypes of early-onset Alzheimer's disease.

OBJECTIVES: We aimed to define brain iron distribution patterns in subtypes of early-onset Alzheimer's disease (EOAD) by the use of quantitative susceptibility mapping (QSM). METHODS: EOAD patients prospectively underwent MRI on a 3-T scanner and concomitant clinical and neuropsychological evaluation, between 2016 and 2019. An age-matched control group was constituted of cognitively healthy participants at risk of developing AD. Volumetry of the hippocampus and cerebral cortex was performed on 3DT1 images. EOAD subtypes were defined according to the hippocampal to cortical volume ratio (HV:CTV). Limbic-predominant atrophy (LPMRI) is referred to HV:CTV ratios below the 25th percentile, hippocampal-sparing (HpSpMRI) above the 75th percentile, and typical-AD between the 25th and 75th percentile. Brain iron was estimated using QSM. QSM analyses were made voxel-wise and in 7 regions of interest within deep gray nuclei and limbic structures. Iron distribution in EOAD subtypes and controls was compared using an ANOVA. RESULTS: Sixty-eight EOAD patients and 43 controls were evaluated. QSM values were significantly higher in deep gray nuclei (p < 0.001) and limbic structures (p = 0.04) of EOAD patients compared to controls. Among EOAD subtypes, HpSpMRI had the highest QSM values in deep gray nuclei (p < 0.001) whereas the highest QSM values in limbic structures were observed in LPMRI (p = 0.005). QSM in deep gray nuclei had an AUC = 0.92 in discriminating HpSpMRI and controls. CONCLUSIONS: In early-onset Alzheimer's disease patients, we observed significant variations of iron distribution reflecting the pattern of brain atrophy. Iron overload in deep gray nuclei could help to identify patients with atypical presentation of Alzheimer's disease. KEY POINTS: • In early-onset AD patients, QSM indicated a significant brain iron overload in comparison with age-matched controls. • Iron load in limbic structures was higher in participants with limbic-predominant subtype. • Iron load in deep nuclei was more important in participants with hippocampal-sparing subtype.

Trial of Deferiprone in Parkinson's Disease.

BACKGROUND: Iron content is increased in the substantia nigra of persons with Parkinson's disease and may contribute to the pathophysiology of the disorder. Early research suggests that the iron chelator deferiprone can reduce nigrostriatal iron content in persons with Parkinson's disease, but its effects on disease progression are unclear. METHODS: We conducted a multicenter, phase 2, randomized, double-blind trial involving participants with newly diagnosed Parkinson's disease who had never received levodopa. Participants were assigned (in a 1:1 ratio) to receive oral deferiprone at a dose of 15 mg per kilogram of body weight twice daily or matched placebo for 36 weeks. Dopaminergic therapy was withheld unless deemed necessary for symptom control. The primary outcome was the change in the total score on the Movement Disorder Society-sponsored revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS; range, 0 to 260, with higher scores indicating more severe impairment) at 36 weeks. Secondary and exploratory clinical outcomes at up to 40 weeks included measures of motor and nonmotor disability. Brain iron content measured with the use of magnetic resonance imaging was also an exploratory outcome. RESULTS: A total of 372 participants were enrolled; 186 were assigned to receive deferiprone and 186 to receive placebo. Progression of symptoms led to the initiation of dopaminergic therapy in 22.0% of the participants in the deferiprone group and 2.7% of those in the placebo group. The mean MDS-UPDRS total score at baseline was 34.3 in the deferiprone group and 33.2 in the placebo group and increased (worsened) by 15.6 points and 6.3 points, respectively (difference, 9.3 points; 95% confidence interval, 6.3 to 12.2; P<0.001). Nigrostriatal iron content decreased more in the deferiprone group than in the placebo group. The main serious adverse events with deferiprone were agranulocytosis in 2 participants and neutropenia in 3 participants. CONCLUSIONS: In participants with early Parkinson's disease who had never received levodopa and in whom treatment with dopaminergic medications was not planned, deferiprone was associated with worse scores in measures of parkinsonism than those with placebo over a period of 36 weeks. (Funded by the European Union Horizon 2020 program; FAIRPARK-II ClinicalTrials.gov number, NCT02655315.).

Optic Nerve Lesion Length at the Acute Phase of Optic Neuritis Is Predictive of Retinal Neuronal Loss.

BACKGROUND AND OBJECTIVES: Acute optic neuritis (ON) is a classical presenting symptom of multiple sclerosis (MS), neuromyelitis optica spectrum disorders (NMOSD), and anti-MOG-associated disorders. The resulting visual impairment is variable and can be severe. Clinicians are in need of predictive biomarkers to optimize the management of acute ON. In this longitudinal study (IRMANO, NCT03651662), we evaluated the ability of optic nerve lesion length measured on MRI at the acute phase of ON to predict retinal neuro-axonal loss and visual impairment at a chronic stage. METHODS: We conducted a longitudinal study (IRMANO, NCT03651662) of patients who presented a clinical episode of ON (≤8 weeks). All patients underwent a retinal optical coherence tomography (OCT) and a brain/optic nerve MRI, including 3D double-inversion recovery (DIR) sequence at the acute phase of ON and 12 months later. Primary outcomes were optic nerve DIR hypersignal lesion length, macular ganglion cell-inner plexiform layer (GCIPL) volume measured on OCT, and low-contrast monocular visual acuity (LCMVA). RESULTS: The study group included 51 patients (33 women, mean age of 32.4 years ± 7.9). We recruited patients with a clinically isolated syndrome (n = 20), a relapsing-remitting MS (n = 23), an isolated ON (n = 6), and a first clinical episode of NMOSD (n = 2). Optic nerve DIR hypersignal was observed in all but 1 symptomatic optic nerves. At inclusion, the mean optic nerve lesion length (in mm) was 12.35 ± 5.98. The mean GCIPL volume (in mm3) significantly decreased between inclusion (1.90 ± 0.18) and M12 (1.67 ± 0.21; p < 0.0001). Optic nerve lesion length at inclusion was significantly associated with GCIPL thinning (estimate ± SD; -0.012 ± 0.004; p = 0.0016) and LCMVA at M12 (0.016 ± 0.003; p < 0.001). Optic nerve lesion length significantly increased at M12 (15.76 ± 8.70; p = 0.0007). The increase in optic nerve lesion length was significantly associated with the GCIPL thinning between inclusion and M12 (-0.012 ± 0.003; p = 0.0011). DISCUSSION: At the acute phase of ON, optic nerve lesion length is an imaging biomarker predictive of retinal neuro-axonal loss and chronic visual impairment, which can help to stratify future therapeutic strategies in acute ON. CLASSIFICATION OF EVIDENCE: This study provides Class I evidence that optic nerve lesion length measured on MRI during the acute phase of a first episode of ON is associated with long-term retinal neuro-axonal loss and visual impairment.

Cerebral mucormycosis: neuroimaging findings and histopathological correlation.

INTRODUCTION: Mucormycosis are infections caused by molds of the order Mucorales. These opportunistic infections are rare, difficult to diagnose, and have a poor prognosis. We aimed to describe common radiographic patterns that may help to diagnose cerebral mucormycosis and search for histopathological correlations with imaging data. METHODS: We studied the radiological findings (CT and MRI) of 18 patients with cerebral mucormycosis and four patients' histopathological findings. RESULTS: All patients were immunocompromised and/or diabetic. The type of lesions depended on the infection's dissemination pathway. Hematogenous dissemination lesions were most frequently abscesses (59 lesions), cortical, cortical-subcortical, or in the basal ganglia, with a halo aspect on DWI for lesions larger than 1.6 cm. Only seven lesions were enhanced after contrast injection, with different presentations depending on patients' immune status. Ischemia and hemorrhagic areas were also seen. Vascular lesions were represented by stenosis and thrombosis. Direct posterior extension lesions were bi-fronto basal hypodensities on CT and restricted diffusion without enhancement on MRI. A particular extension, perineural spread, was seen along the trigeminal nerve. Histopathological analysis found endovascular lesions with destruction of vessel walls by Mucorales, microbleeds around vessels, as well as acute and chronic inflammation. CONCLUSIONS: MRI is the critical exam for cerebral mucormycosis. Weak ring enhancement and reduced halo diffusion suggest the diagnosis of fungal infections. Involvement of the frontal lobes should raise suspicion of mucormycosis (along with aspergillosis). The perineural spread can be considered a more specific extension pathway of mucormycosis.

Longitudinal Analysis of Brain-Predicted Age in Amnestic and Non-amnestic Sporadic Early-Onset Alzheimer's Disease.

Objective: Predicted age difference (PAD) is a score computed by subtracting chronological age from "brain" age, which is estimated using neuroimaging data. The goal of this study was to evaluate the PAD as a marker of phenotypic heterogeneity and severity among early-onset Alzheimer's disease (EOAD) patients. Methods: We first used 3D T1-weighted (3D-T1) magnetic resonance images (MRI) of 3,227 healthy subjects aged between 18 and 85 years to train, optimize, and evaluate the brain age model. A total of 123 participants who met the criteria for early-onset (<65 years) sporadic form of probable Alzheimer's disease (AD) and presented with two distinctive clinical presentations [an amnestic form (n = 74) and a non-amnestic form (n = 49)] were included at baseline and followed-up for a maximum period of 4 years. All the participants underwent a work-up at baseline and every year during the follow-up period, which included clinical examination, neuropsychological testing and genotyping, and structural MRI. In addition, cerebrospinal fluid biomarker assay was recorded at baseline. PAD score was calculated by applying brain age model to 3D-T1 images of the EOAD patients and healthy controls, who were matched based on age and sex. At baseline, between-group differences for neuropsychological and PAD scores were assessed using linear models. Regarding longitudinal analysis of neuropsychological and PAD scores, differences between amnestic and non-amnestic participants were analyzed using linear mixed-effects modeling. Results: PAD score was significantly higher for non-amnestic patients (2.35 ± 0.91) when compared to amnestic patients (2.09 ± 0.74) and controls (0.00 ± 1). Moreover, PAD score was linearly correlated with the Mini-Mental State Examination (MMSE) and the Clinical Dementia Rating Sum of Boxes (CDR-SB), for both amnestic and non-amnestic sporadic forms. Longitudinal analyses showed that the gradual development of the disease in patients was accompanied by a significant increase in PAD score over time, for both amnestic and non-amnestic patients. Conclusion: PAD score was able to separate amnestic and non-amnestic sporadic forms. Regardless of the clinical presentation, as PAD score was a way of quantifying an early brain age acceleration, it was an appropriate method to detect the development of AD and follow the evolution of the disease as a marker of severity as MMSE and CDR-SB.

Prediction of Long-term Cognitive Functions after Minor Stroke, Using Functional Connectivity.

OBJECTIVE: To determine whether functional MRI connectivity can predict the long-term cognitive functions 36 months after minor stroke. METHODS: Seventy-two participants with first-ever stroke were included at baseline and followed up for 36 months. A ridge regression machine learning algorithm was developed and used to predict cognitive scores 36 months post-stroke on the basis of the functional networks measured using MRI at 6 months (referred to here as the post-stroke cognitive impairment (PSCI) network). The prediction accuracy was evaluated in four domains (memory, attention/executive, language and visuospatial functions) and compared with clinical data and other functional networks. The models' statistical significance was probed with permutation tests. The potential involvement of cortical atrophy was assessed 6 months post-stroke. A second, independent dataset (n=40) was used to validate the results and assess their generalizability. RESULTS: Based on the PSCI network, a machine learning model was able to predict memory, attention, visuospatial functions and language functions 36 months post-stroke (r2: 0.67, 0.73, 0.55 and 0.48, respectively). The PSCI-based model was at least as accurate as models based on other functional networks or clinical data. Specific patterns were demonstrated for the four cognitive domains, with involvement of the left superior frontal cortex for memory, attention and visuospatial functions. The cortical thickness 6 months post-stroke was not correlated with cognitive function 36 months post-stroke. The independent validation dataset gave similar results. CONCLUSIONS: A machine learning model based on the PSCI network can predict the long-term cognitive outcome after stroke.

Intraoperative MRI and FLAIR Analysis: Implications for low-grade glioma surgery.

PURPOSE: Intraoperative MRI (iMRI) offers the possibility of acquiring intraoperatively real-time images that will guide neurosurgeons when removing brain tumors. The objective of this study was to report the existence of FLAIR abnormalities on iMRI that may occur on the margin of a brain resection and may lead to misdiagnosis of residual tumor. METHODS: We retrospectively analyzed intraoperative MRI (iMRI) in 21 consecutive patients who underwent surgery for a low-grade glioma. Two readers independently reviewed iMRI images to search for the presence of a FLAIR hyperintensity surrounding the surgical cavity. For each patient, they were instructed to characterize FLAIR abnormalities on the margins of the resected area as (1) no FLAIR abnormality; (2) "linear FLAIR hyperintensity (LFH)", when a<5mm linear FLAIR hyperintensity was present; or (3) "nodular FLAIR hyperintensity (NFH)", in the case of a thick and nodular FLAIR hyperintensity. RESULTS: LFH were present on at least one surgical margin of one third of the patients analyzed with iMRI, and vanished on follow-up MRI, confirming its transient condition; whereas NFH were linked to persistence of pre-surgical abnormalities, such as residual tumor as confirmed or by histopathological analysis of a second surgery or by its remnant on follow-up MRI. CONCLUSION: Linear FLAIR hyperintensities can be present on surgical margins analyzed by iMRI and should not be mistaken for residual tumor.