Using Cerebrospinal Fluid Improves Detection of Individual Brain Atrophy

Using Cerebrospinal Fluid Improves Detection of Individual Brain Atrophy | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Using Cerebrospinal Fluid Improves Detection of Individual Brain Atrophy Calvin W. Howard, Sheena Barotono, Elmira Hassanzadeh, Grace Burt, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6423862/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Clinical neuroradiologists routinely look for expansion of CSF spaces to help identify atrophy on patient MRI scans. In contrast, automated methods for identifying atrophy rely on changes in grey matter volume or cortical thickness. It is unclear if evaluating CSF spaces could improve detection of brain atrophy, which may be relevant to improving detection of age- and disease-related atrophy. Methods 3 clinician experts graded atrophy across 7 brain regions from 50 subjects enrolled in the Alzheimer Disease Neuroimaging Initiative (Discovery Cohort, n = 1050 visual ratings) while one expert graded atrophy in an additional 150 subjects (Validation cohort, n = 1050 visual ratings). These subjects included patients with mild cognitive impairment (MCI, n = 72), Alzheimer’s disease (AD, n = 60), and age-matched healthy controls (n = 68), randomly selected from the broader sample. We used an automated approach to detect expansion of CSF spaces and compared it with standard methods for detecting brain atrophy (grey/white matter volume, cortical thickness). We evaluated four metrics of performance: 1) correlation to visually rated atrophy; 2) correlation to clinical symptoms; 3) localization of atrophy most correlated with verbal memory scores; and 4) ability to discriminate between AD, MCI, and controls. Results Atrophy detected by expansion of CSF spaces significantly outperformed existing methods across all performance metrics. 1) CSF-based atrophy correlated with manually assessed atrophy scores (Median Rho = 0.50, p max = 0.043), and this correlation was stronger than all other methods (max p FWE = 0.0005). 2) CSF-based atrophy correlated with clinical symptoms (Median Rho = 0.37, IQR 0.34–0.46), and this correlation was stronger than all other methods (max p FWE = 0.0015). 3) CSF-based atrophy was the only method to localize FWE-significant atrophy covarying with verbal memory scores to the left hippocampus (Rho = 0.57, p FWE = 0.00302). 4) CSF-based atrophy best differentiated between AD, MCI, and controls (AUC = 0.68, 95% CI 0.61–0.75), and outperformed all other methods (max p FWE = 0.041). All results were reproducible across discovery and replication cohorts. Conclusion Deriving brain atrophy using CSF can increase sensitivity of atrophy detection, improving alignment with clinical evaluations, explained variance, localization strength, and diagnostic utility. Biological sciences/Neuroscience Health sciences/Anatomy/Nervous system Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction When neurologists and neuroradiologists are working up a dementia diagnosis, they routinely assess patient brain scans for disease-specific atrophy patterns. 1 In clinical practice, these atrophy patterns are often identified qualitatively, 2 by visually inspecting the contrast between brain and cerebrospinal fluid (CSF) for the hallmark widening of CSF that occurs with brain atrophy. 3 , 4 Despite this qualitative nature, clinical inspection of CSF to proxy brain atrophy both accurately identifies atrophy location 5 and neurodegenerative diagnosis. 4 Although it may be useful clinically, using widened CSF spaces to detect atrophy is only a proxy of brain atrophy, not a direct measure of neuronal loss. These widened CSF spaces could represent multiple processes, such as atrophy, increased CSF pressure, or increased glymphatic space. 4 , 6 , 7 As such, most atrophy detection software and research has focused on more direct measures of grey matter (GM) volume, white matter (WM) volume, cortical thickness (CTh). 5 , 8 , 9 Despite evaluating the brain differently than clinicians, these methods robustly map brain atrophy, 1 , 10 , 11 explain symptom variance, 4 , 12 , 13 and aid diagnosis. 14 – 16 Given the established strengths of GM, WM, and CTh, only a handful of research studies have explored atrophy detection software using widened CSF spaces as a proxy for atrophy. 17 – 21 In these studies, CSF performed surprisingly well in atrophy detection, 17 – 21 disease diagnosis, 14 and sometimes outperformed assessments based on GM, WM, and CTh. 14 , 17 – 19 While these findings are interesting, most of these results required specialized MRI sequences, 20 , 22 were restricted to specific brain regions, 14 or focused on group-level findings rather than individual patients. 14 , 17 , 19 Thus, the utility of CSF for assessment of a patient’s atrophy remains unclear. Here, we evaluate an automated approach to map brain atrophy in individual patients based on expansion of CSF spaces visible on routine clinical MRI. By comparing a patient’s CSF to a reference ‘healthy’ distribution, we generate a patient-specific map of CSF-based atrophy as a proxy of the patient’s underlying atrophy. We then tested the clinical relevance of this map using four outcome metrics: 1) correlation with clinician evaluations, 2) correlation with clinical symptoms, 3) localization potential, and 4) diagnostic potential. For all outcome metrics, we compared results to conventional atrophy detection algorithms based on GM, WM, and CTh. Methods Ethics Statement This study was conducted in accordance with ethical standards and approved by the Institutional Review Board of the Brigham and Women’s Hospital and Harvard Medical School, Boston, Massachusetts. Given the secondary use of research data, the study was exempted from obtaining informed consent. Data used in this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). 23 Institutional Review Board (IRB) approval was obtained from all sites, and informed consent obtained from all participants. Informed written consent was collected from all subjects within each site. ADNI provided access to all data used in this study and the authors used their guidelines for acknowledging them in the author section. Subjects All subjects were recruited from ADNI1 (2004–2009). 23 Each patient received an initial visit with a physical examination, neurological examination, neuropsychological examination, fluid biomarkers, and neuroimaging. 23 At the end of this visit, they received a diagnosis of cognitively normal (control), mild cognitive impairment (MCI), or Alzheimer Disease (AD). Diagnosis was based on clinical criteria demonstrating a combination of Mini Mental Status Examination Score 24 , Clinical Dementia Rating score 25 , Wechsler Memory Scale-Revised 26 , with additional criteria for: dementia 27 , MCI 28 , and AD 29 (Supplementary Table 1). A detailed breakdown of inclusion/exclusion criteria used in ADNI 23 is available (Supplementary Table 2). Overall, 338 subjects were included from ADNI (74.0 ± 6.8 years, 48.5% female). To compare patient brains against a reference distribution, we collected a reference cohort of ‘healthy’ controls (n = 138, 73.7 ± 6.4 years, 48.5% female). We also collected a separate cohort of patients for evaluation (n = 200, 73.0 ± 7.0 years, 47.5% female), composed of controls (n = 68, 72.4 ± 7.1 years, 42.9% female), MCI patients (n = 72, 72.4 ± 7.1 years, 42.9% female), and AD patients (n = 60, 74.5 ± 6.9 years, 54% female). The patient cohorts were randomly subsampled from ADNI in a 1:1:1 ratio. We then split the 200 patients into discovery (n = 50, for initially assessing the algorithm’s utility) and validation cohorts (n = 150, for testing the algorithm in unseen patients). The discovery cohort (n = 50, 73.9 ± 7.8 years, 36.7% female) was composed of control (n = 21), MCI (n = 20), and AD patients (n = 9). The validation cohort (n = 150, 74.8 ± 6.9 years, 51.3% female) was composed of control (n = 47), MCI (n = 52), and AD patients (n = 51). Demographic characteristics of this cohort are available (Supplementary Table 3). Neuroimaging The MRI protocol for ADNI1 (2004–2009) acquired structural imaging on 3T scanners using T1 sequences (MPRAGE protocol: sagittal plane, TR/TE/TI, 2400/3/1000 ms, flip angle 8°, 24 cm FOV, 192 × 192 in-plane matrix, 1.2 mm slice thickness). 30 In this study, we evaluated only T1 scans acquired on 3T scanners. Clinician Evaluation of Atrophy A neuroradiologist (EH), a cognitive neurologist (SB), and a clinical neuroimaging expert (CH) assessed 2100 brain regions across 200 patient brain scans, spanning 1050 regions in the 50 discovery cohort patients (EH, SB, CH assessed 7 regions across 50 patients), and 1050 regions in the 150 validation cohort patients (CH assessed 7 regions across 150 patients). All clinicians were blinded, rated independently of each other, and instructed in use of the appropriate visual rating scale. EH, SB, and CH evaluated the validation cohort. CH evaluated the validation cohort. All clinician evaluations of atrophy occurred before quantitative atrophy measurements. No clinical information was provided (i.e. age), ensuring raters evaluated the brain scans in isolation. Visual Rating Scales In all patients, seven regions were evaluated: the frontal lobe, temporal lobe, parietal lobe, occipital lobe, mesial temporal lobe, and cerebellum; ventriculomegaly was also graded. In cases of asymmetric atrophy, the worse grade was taken. The mesial temporal lobe was graded using the Mesial Temporal Atrophy (MTA) scale, 31 the most reliable and widely used visual scale for mesial temporal atrophy. 32 , 33 The MTA scale ranges from 0 (intact) to 4 (severe). The full MTA scale is available in the supplements (Supplementary Table 4). The other 6 out of 7 regions were evaluated with the Global Cortical Atrophy (GCA) scale, 34 as it is the most widely used assessment scale of brain-wide atrophy. 2 , 3 , 35 was applied to all other regions. Each region was evaluated independently of other regions, as has been previously done. 3 The GCA scale ranges from 1 (intact) to 4 (severe). The full GCA scale is available in the supplements (Supplementary Table 5). Reliability of Clinician Evaluations The interrater reliability among the 3 clinicians was evaluated across all regions using intraclass correlation coefficient (ICC) and comparison of atrophy rating distributions via Kolmogorov Smirnov Test. Author CH was compared specifically to the neuroradiologist (author EH) to ensure validity of ratings of the validation cohort. Voxel Based Morphometry VBM analyses were performed with Statistical Parametric Mapping (SPM) and Computational Anatomy Toolbox (CAT12) to derive maps of each patient’s GM, WM, and CSF segments as previously described. 9 , 36 Briefly, all 3D T1-weighted MRI scans are preprocessed with spatial adaptive non-local means (SANLM) denoising filter. 37 These are then bias corrected and affine registered to the space of SPM’s tissue probability maps. 9 Next, GM, WM, and CSF components are segmented out using the SPM tissue probability maps, 9 resulting in probability maps for each segment. 38 Diffeomorphic Anatomic Registration Through Exponentiated Lie Algebra (DARTEL) is used to normalize the segmented scans into MNI space. 39 , 40 DARTEL provides precise normalization to MNI space, and has the advantage of preserving structural anomalies inherent to disease processes. 40 , 41 This generates a map of ‘concentration’ of each segment of the brain across all voxels in MNI space. The Jacobian determinant derived from the warp is then applied to the voxels, modulating them and converting them to an adjusted volume which accounts for the distortion by the warp process. 9 The resulting tissue segment maps were smoothed with both 0mm and 6mm isotropic smoothing kernels. This process results in the final grey, white, and cerebrospinal fluid volume maps. Surface Based Morphometry SBM analyses were performed with FreeSurfer 5.3.0 to derive CTh values across each patient’s brain. 8 In brief, each patient’s brain was affine-registered to MNI space. Subsequently, WM segmentation is performed to delineate the edge of the WM, which is then tessellated into a surface. The outer boundary of the cortex is then estimated by identifying the CSF edge. The radial distance between the WM surface and the outer boundary of the cortex is then estimated at all points, representing the thickness of the cortex (CTh). Mapping Patient Atrophy There are two established methods for defining atrophy in reference to a control distribution: Z-scoring 42 – 46 and W-scoring 10 , 47 , 48 . Briefly, we calculate voxelwise Z- and W-scores by defining a normal distribution based on the 138 healthy control subjects, then evaluating the voxelwise deviation in experimental subjects. Z-scoring simply compares each patient’s brain to the expected mean, and normalizes the difference based on standard deviation (Supplementary Eq. 1). W-scoring fits a linear model to the control distribution, then uses that to predict the expected value of the patient and normalizes this by the residual standard deviation (Supplementary Eq. 2). We derived both for all patients, but present Z-scores in the primary results as they were ultimately found to be superior across all evaluation metrics. The Z- or W-maps can be thresholded at a value of 2, representing volume loss beyond the 98th percentile of expected. This is our operational definition of atrophy. GM, WM, and CT are thresholded below a value of -2 to identify atrophy, while CSF is thresholded above + 2. Composite atrophy maps were created by summating thresholded atrophy maps. All combinations of composite maps across each method (CSF, GM, WM, CTh) were created. We generated atrophy maps for all patients across smoothing kernels of 0mm and 6mm. Reference controls were smoothed to 6mm prior to comparing patients against them. To map expansion of CSF spaces, we compared the CSF in each patient to a normative distribution of CSF from healthy controls (Supplementary Fig. 8). The result is a voxel-wise map, where values represent deviations in the patient relative to the normal distribution (Fig. 1 ). Note that this map of CSF-based atrophy is only an indirect proxy of underlying brain atrophy, we will refer to it as a “CSF-based atrophy map” to match the terminology used for other tissue components “e.g. GM-based atrophy map”. Measuring Degree of Atrophy in Each Brain Region After deriving either Z- or W-maps of each patient’s atrophy, we evaluated atrophy in all 7 regions of interest (mesial temporal, frontal, temporal, parietal, and occipital lobes, as well as cerebellum, and periventricular subcortex). Z- or W-scores within each region were summated. This provided an evaluation of quantitative atrophy in each region of interest. Relating Atrophy to Clinician Assessments Given the ordinal nature of the visual rating scales, Spearman Correlation related quantified atrophy to clinician atrophy scores. For every lobe, across each method, we correlated atrophy to clinician atrophy scores. Kruskal-Wallis tests compared performance of each method, using performance across the 7 ROIs to estimate the distribution of the method’s overall performance. Post-hoc testing used permutation analysis to compare significant difference between correlations across methods. ANOVA and permutation testing evaluated differences within each region, across methods. Comparisons were made with and without the ventricles and subcortex regions of interest to ensure fair comparison to CTh. All evaluations presented in the figures are relative to the clinical neuroradiologist and robust when compared to the neurologist or resident. Relating Atrophy to Cognitive Outcomes To test whether atrophy was related to clinical symptoms, we examined the relationship between regional brain atrophy and each patient’s baseline Alzheimer’s Disease Assessment Scale of Cognition-11 (ADAS-Cog 11). The ADAS-Cog 11 was chosen as it is the most comprehensive metric of cognitive function in Alzheimer’s, and is ubiquitously available in ADNI. 49 Given the ordinal nature of the ADAS-Cog 11, Spearman correlations related atrophy to each patient’s ADAS-Cog 11. Kruskal-Wallis tests compared performance of each method, using performance across the 7 ROIs to estimate the distribution of the method’s overall performance. Post-hoc testing used permutation analysis to compare significant difference between correlations across methods. ANOVA and permutation testing evaluated differences within each region, across methods. Comparisons were made with and without the ventricles and subcortex regions of interest to ensure fair comparison to CTh. Using Atrophy to Localize Delayed Word Recall We assessed if the atrophy methods could localize a neurological function that was both highly focal and known to be affected in neurodegeneration. We chose delayed word recall (ADAS-Cog 11 question 4), a function of memory impaired by Alzheimer’s, which is known to localize primarily to the left hippocampus by histopathological study 50 – 54 The ability for each method to localize delayed word recall was measured via whole-brain data-driven analysis and in an a priori region of interest analysis. The whole-brain analysis used voxel- or vertex-wise Spearman Correlations to relate atrophy at each point to a patient’s verbal recall. Permutation testing shuffled outcomes to estimate p-value, which were maximum statistic family-wise error corrected. 55 The difference in peak hippocampal correlations were also compared across methods using permutation analysis. A hippocampal region of interest was used to quantify atrophy for each patient. Atrophy within this hippocampal region was then correlated with each patient’s delayed word recall. Permutation testing compared the strength of this correlation across methods. Using Atrophy to Diagnose Patients as Control, MCI, or AD We also assessed how well each method could detect diagnostically relevant atrophy. To do this, we employed a standard method of discriminating between patients: by calculating atrophy within regions of interest and using it to fit a logistic regression on diagnosis. 15 Atrophy within the frontal, parietal, temporal, occipital, and mesial temporal lobes were quantified for each disease. Then, a multiclass logistic regression related atrophy in these regions to classify patients as control, MCI, or AD. The logistic regressions were fit on our well-characterized discovery dataset that all clinicians agreed upon (n = 50) and were subsequently tested on the held-out validation dataset (n = 150). Performance was measured by each model’s area under the receiver operating characteristic (AUROC). Bootstrapping (n = 10 000) was used to estimate confidence intervals for each AUROC. A smaller training dataset than validation dataset is expected to be more rigorous, and is known to produce lower AUROC estimates. 56 Significant differences between bootstrapped confidence intervals were evaluated using two methods: 1) DeJong’s test, 57 and 2) proportion of superior samples. 58 Statistics Statement Statistical analyses were conducted in Python 3.10 using Statsmodels 0.14.0, Nilearn 0.10.1, and Scikit-learn 1.3.0. 59 – 61 Descriptive statistics are presented as mean ± standard deviation. Correlations used Spearman's methods as appropriate for ordinal data. Logistic regressions were performed with multiclass multinomial logistic regressions. Comparison of atrophy distributions was performed with Kolmogorov-Smirnov tests with Bonferroni family wise error correction. Permutation testing was performed as previously described. 55 Briefly, a statistic of interest (Spearman R) was calculated on the actual data. Then, the outcomes were shuffled and the statistic was recomputed; this occurred multiple times (n = 10 000). The number of times the random process exceeded the observed value, averaged, provides the p-value. Results Clinician Evaluation of Patient Atrophy Three clinical raters scored atrophy across 7 brain regions using standard visual rating scales for 50 subjects in our discovery cohort (n = 1050 atrophy scores). Atrophy ratings between a board-certified neuroradiologist and two independent clinical experts were reliable (Supplementary Fig. 1) across all brain regions (mean ICC = 0.81, max p FWE = 0.0036). One clinical expert analyzed the additional 150 validation subjects (n = 1050 atrophy scores). This expert’s evaluations were validated against the neuroradiologist using ICC (Supplementary Fig. 2) and reliably rated atrophy compared to the neuroradiologist (mean ICC = 0.82, max p FWE = 0.0005). We next evaluated if the experts identified the standard distribution of atrophy in AD and MCI patients. AD patients had more atrophy than controls in the frontal, temporal, parietal, and mesial temporal lobes (min U = 11, max p FWE = 0.0063). MCI patients had more atrophy than controls in the frontal and temporal lobes (min U = 8.2, max p FWE = 0.041). Rating distributions were not found to be significantly different across raters (p Min > 0.05, Supplementary Fig. 3). The distribution of atrophy in the validation cohort was not significantly different from the discovery cohort, regardless of which rater was used for comparison (min p = 0.33). Automated Evaluation of Individual Atrophy We next generated maps of whole-brain atrophy for each patient based on GM, WM, CTh, and CSF. The atrophy maps often failed to align both with each other and with the clinical ratings, although CSF did align with clinical evaluations (Fig. 2 ). GM, WM, and CTh-based methods also detected locations of brain hypertrophy rather than atrophy, although this hypertrophy was not detected by the clinicians nor CSF (Supplementary Fig. 4). We next averaged all atrophy maps across 50 AD patients and counted the number of hypertrophic voxels identified. Less than 1% of voxels evaluated with CSF were found to be hypertrophic in these 50 AD patients, compared to 37% in GM, 22% in WM, and 5% in CTh (Supplementary Fig. 5). We next investigated if processing errors were contributing to the lower performance in GM, WM, and CTh. As a control, we investigated if these segments could recompose the distribution of AD-related atrophy they have previously defined at the group level. We summed and averaged individual atrophy across 50 AD patients, derived by GM, WM, and CTh, and found that these did recomposed the expected precuneus and mesial temporal dominant atrophy with associated frontal and lateral temporal (Supplementary Fig. 4). 45 , 46 We next investigated what might be causing higher performance with CSF. At the level of the voxel, CSF improved sensitivity to outlier voxels, which we show in relation to our example patient’s whole-brain peak atrophy (Supplementary Fig. 7). At the level of the brain, CSF improved whole-brain coverage and enabled detection of missing brain regardless of whether the source of atrophy was from GM, WM, or cortex (Supplementary Fig. 8). Atrophy Detected by CSF Correlates best with Clinician Ratings We correlated the detected atrophy in each brain region to the atrophy rating from the neuroradiologist (Fig. 3 A). Across methods, there was a difference in strength of correlation to the neuroradiologist (Kruskal-Wallis, p = 0.0087). Post-hoc permutation testing demonstrated CSF showed the best correlation with clinical ratings (Rho median = 0.50, IQR 0.31–0.79) and was more correlated to clinician ratings than GM (DRho = 0.42, p < 0.0001), WM (DRho = 0.41, p < 0.0001), and CTh (DRho = 0.30, p = 0.0005). We repeated the above analyses but assessed each of the 7 brain regions separately (Fig. 3 B). Correlations to clinician-detected atrophy again varied significantly across CSF, GM, WM, and CTh (ANOVA, p FWE Max = 0.0091). Atrophy was significantly correlated with the neuroradiologist rating in for 7/7 regions with CSF (p max = 0.043), 0/7 regions with GM (p max = 0.72), 0/7 regions with WM (p max = 0.91), and 2/7 regions with CTh (p max = 0.87). Results were not driven by specific diagnostic subgroups in the discovery cohort (Supplementary Fig. 9), controls (Supplementary Fig. 10), and were robust across different raters. The analysis was repeated in the full set of 150 test patients, 50 control patients, 50 MCI patients, and 50 AD patients with consistent results (Supplementary Fig. 11). We next compared different atrophy processing pipelines to understand how CSF performs across them. We found CSF outperformed other methods across various smoothing kernels, but using W-scoring instead of Z-scoring to define atrophy eliminated the performance of CSF (Supplementary Fig. 12) and reduced the performance of each atrophy method to below-chance level (Supplementary Fig. 13). We tested the differences between Z- and W-scores, finding the difference was not due to controlling for covariates (Supplementary Fig. 14), isolating the difference as the model-free nature of Z-scores versus the linear model-based nature of W-scores. Atrophy Detected by CSF Correlates with Cognitive Symptoms We next tested whether atrophy detected by each method was correlated with ADAS-Cog 11 scores (Fig. 4 A). There was a significant difference in this correlation across methods (Kruskal-Wallis, p = 0.0042). CSF-based atrophy was most correlated with cognitive scores (Rho median = 0.37, IQR 0.34–0.46) and more correlated to cognitive scores than GM atrophy (DR = 0.36, p < 0.0001), WM atrophy (DR = 0.30, p < 0.0001), and CTh atrophy (DR = 0.22, p = 0.0066). Results were robust to using the Clinical Dementia Rating scale sum of boxes (Kruskal-Wallis, p = 0.019), Mini Mental Status Exam (Kruskal-Wallis, p = 0.0092,) with or without inclusion of the cerebellum and subcortex, and were replicated in the validation cohort (Kruskal-Wallis, p = 0.0018). We next repeated the correlation of atrophy to symptoms but assessed each of the 7 brain regions independently (Fig. 4 B). The strength of the relationship to clinical symptoms were found to vary significantly across CSF, GM, WM, and CTh (ANOVA, p FWE = 0.0029). Atrophy was significantly correlated with clinical symptoms across 7/7 regions using CSF (p max = 0.043), 1/7 regions using GM (p max = 0.60), 1/7 regions using WM (p max = 0.93), and 1/7 regions using CTh (p max = 0.18). Control analyses were robust to evaluation in the 150 validation patients, comparing correlation strengths nonparametrically (permutation test, p FWE Max = 0.017), and were again optimal with Z-scoring (Supplementary Fig. 15) compared to W-scoring (Supplementary Fig. 16). In additional analyses, we repeated the above analyses but compared each method to the neuroradiologist. Overall, CSF was found to relate more strongly to clinical symptoms than the neuroradiologist (p = 0.012, Supplementary Fig. 17). When assessing each lobe independently, CSF-based atrophy was found to relate more strongly to symptoms than the neuroradiologist’s visual assessment of atrophy in 5/7 lobes (p Max = 0.047), which was unique among the automated methods (Supplementary Fig. 18). Atrophy Detected by CSF can Localizes specific Neurological deficits Next, we wondered if CSF-based atrophy was spatially accurate enough to localize specific neurological deficits. We attempt to localize atrophy impairing delayed word recall, which is known to localize to the left hippocampus. 54 , 62 We used a whole-brain data-driven correlation to relate voxelwise atrophy to delayed word recall performance (Fig. 5 ). This was repeated for each method. CSF was the only method to successfully localize the whole-brain peak (maximum) correlation to the left hippocampus (Rho = 0.57, p FWE = 0.00302). Permutation testing revealed the CSF peak was significantly stronger than the hippocampal peaks localized by GM (p FWE = 0.042), WM (p FWE = 0.006), and CTh (p FWE = 0.004). This hippocampal localization was specific to delayed word recall and was not similarly localized when this analysis was repeated using the other ten ADAS-Cog 11 subscores. CSF was also the only method to successfully localized delayed word recall after controlling for other cognitive subscores (Supplementary Fig. 21) or controlling for age- and sex-related atrophy (Supplementary Fig. 22). We verified the above results using an a priori hippocampus region of interest. For each method, we quantified left hippocampal atrophy and correlated it to delayed word recall (Supplementary Fig. 19). The strongest correlation was seen with CSF atrophy (Rho = 0.59, p = 0.0000), followed by CTh (Rho = 0.47, p = 0.0010), GM (Rho = 0.31, p = 0.039), and WM (Rho = 0.24, p = 0.11). The correlation between CSF atrophy and verbal memory was significantly stronger than GM (DR = 0.28, p FWE = 0.013), WM (DR = 0.35, p FWE = 0.0083), and CTh (DR = 0.12, p FWE = 0.048). This analysis was replicated in the validation cohort, with CSF again showed the strongest correlation with delayed word recall (Rho = 0.68, p = 0.0000), followed by CTh (Rho = 0.42, p = 0.0021), GM (Rho = 0.38, p = 0.027), and WM (Rho = 0.19, p = 0.18). CSF was best able to localize additional ADAS-Cog 11 subscores to a-priori lobar localizations (Supplementary Table 6). Atrophy Detected by CSF Discriminates Between Diagnoses Next, we used the atrophy detected across the frontal, parietal, occipital, temporal, and mesial temporal lobes to distinguish between diseases (Fig. 6 ). A logistic regression was trained on the discovery cohort evaluated by the neuroradiologist (n = 50), then tested in the held-out validation cohort (n = 150). The logistic regression using CSF achieved the highest overall discrimination between control, MCI, and AD patients (AUROC = 0.67, p = 0.0001), which permutation testing demonstrated was significantly stronger than discrimination from GM (p = 0.0081), WM (p = 0.038), and CTh (p = 0.0046). Differences were robust to DeLong’s test (max p FWE = 0.001). CSF was found to be more accurate than the radiologist (p = 0.012), which was unique among the automated methods (Supplementary Section 6.1). We next wondered if the performance of the CSF-based classifier was driven by any specific diagnosis (AD, MCI, or control). We found CSF improved discrimination of AD from MCI patients (AUROC = 0.75, p = 0.0001), which was significantly improved compared to GM (p = 0.0081), WM (p = 0.021), and CTh (p = 0.012). All methods performed well at classifying AD vs Control, as well as MCI vs Control patients, with no differences between the atrophy methods (Supplementary Fig. 23). Combining CSF-based atrophy mapping with other methods We next explored whether CSF-based atrophy mapping could be combined with other atrophy measures to improve performance on any of our four outcome measures. For correlation with neuroradiologist’s scores (Supplementary Fig. 24), adding CSF to other atrophy methods improved correlations (p max = 0.032) but nothing outperformed CSF alone (p min = 0.87). For correlation with cognitive scores (Supplementary Fig. 25), adding CSF to other atrophy methods improved correlations (p = 0.37). Multivariate regression (Supplementary Fig. 26) demonstrated the combined CSF + GM atrophy maps explained the most clinical variance in ADAS-Cog scores (R 2 = 0.54, p = 0.0001), and outperformed CSF alone (DR 2 = 0.15, p FWE = 0.0001). For localization of verbal memory, adding CSF to other atrophy scores improved the correlation with delayed word recall (Rho min = 0.66, p max = 0.0000) but did not outperform correlations based on CSF alone (p min = 0.11). Finaly, for diagnostic discrimination we found CSF combined with GM provided the overall highest discrimination of control, MCI, and AD patients (AUROC = 0.68, p = 0.001), but not significantly better than CSF alone (p = 0.32). Discussion We found CSF-based atrophy was most consistent with clinical evaluations, which is not surprising as clinical evaluation is based in part on visualizing expanded CSF spaces. 3 , 64 However, we also found that CSF-based atrophy was also most correlated with clinical symptoms, best localized neurological functions, and yielded the highest diagnostic utility. These latter findings are surprising, as CSF expansion is only an indirect proxy of neurodegeneration and lacks the neuroanatomical specificity of techniques like GM or CTh. There are good reasons to think that CSF-based metrics would not perform well in detecting clinically relevant brain atrophy, as the majority of neuroimaging research has focused on more direct measures of atrophy based on grey matter, white matter, or cortical thickness. 8 , 9 However, expansion of CSF spaces remains a hallmark of the clinical approach to identifying atrophy on MRI scans. 31 , 34 , 65 Our results support this clinical practice and may provide insight into why this clinical approach to atrophy detection has survived the test of time. Neurodegenerative diseases such as AD are known to impact many tissue types simultaneously, including grey mater, white matter, and cortical thickness. 16 , 45 – 47 , 66 , 67 Techniques focused on just one of these tissue types alone may fail to capture clinically relevant atrophy in the other tissue types. While CSF-based atrophy is non-specific, it may do a good job of capturing the summed effect of atrophy across tissue types. Second, the boundary between CSF and non-CSF is easier to detect due to high contrast between CSF and brain. In contrast, it may be harder to detect boundaries between brain tissue types (e.g. GM vs WM). This may render visual or automated methods based on CSF more accurate. Future work is needed to determine why CSF performs well. Relevance to Aging Atrophy can be caused by a number of etiologies, with common etiologies being neurodegeneration and aging. In this manuscript, we compared two different methods of defining brain atrophy, the z-score and w-score. We found the Z-score provides the best overall assessment of atrophy, and this has one critical implication for aging. Z-scores inherently detect age-related atrophy, while W-scores remove age-related atrophy. Thus, CSF-based z-scored atrophy may be a useful tool for aging researchers to better understand how age influences the brain, in derivation of brain-age metrics, or as a useful feature in machine learning algorithms. Relevance to Research Automated atrophy detection methods using GM, WM, and CTh work well and have a longstanding history of successful applications to structural imaging. 1 , 4 , 10 – 16 However, we found that CSF added additional performance beyond these more common methods, a finding which is consistent with a small number of studies that have examined CSF-based atrophy. Our results confirm and expand this prior work, showing that CSF-based methods outperform other GM, WM, and CTh across 4 different clinically relevant metrics. We also found that CSF-based atrophy might be combined with other techniques to provide additional explanatory power. Our code is openly available ( https://github.com/Calvinwhow/vbm.git ). Relevance to Clinical Practice CSF-based atrophy correlated well with atrophy assess by a board-certified clinical neuro-radiologist, which may prove useful in ongoing attempts to complement clinical radiology with automated methods. 17 – 21 In our study, CSF-based atrophy correlated better with clinical symptoms than atrophy assess by the radiologist. This suggests that our automated method for detecting CSF-based atrophy may detects clinically-relevant atrophy which is not obvious on visual inspection. Our results suggest that CSF-based atrophy adds additional discriminatory value in distinguishing between AD, MCI, and control patients. This added value was specific to separating MCI from AD, a task which has traditionally been more difficult compared to discriminating patients with AD or MCI from controls. 3 , 33 , 33 , 43 Limitations Several limitations apply to this work. First, CSF may not detect atrophy from regions deep within the brain that are not adjacent to CSF spaces. Whether CSF-based atrophy performs well in other disorders outside of MCI/AD, especially those associated with more subcortical atrophy, remains unknown. Second, we performed our neuroimaging analyses in atlas space, which has the potential for registration inaccuracies, especially in the context of brain atrophy. 68 , 69 We used a registration algorithm specifically designed to minimize atrophy-related distortions 40 but future work is needed to determine the ideal co-registration algorithm or whether automated methods in patient space may perform better. A third limitation is our evaluation of the cerebellum. While the CSF-based atrophy correlated with clinician assessments of cerebellar atrophy, we note the clinicians only had ‘good’ intraclass correlation in cerebellar evaluation. In part, this is driven by the lack of an existing visual cerebellar atrophy rating scale to act as a gold-standard. Future work will need to develop a gold-standard cerebellar rating scale and relate quantified atrophy to clinician ratings. Declarations Author Contribution C.W.H. conceptualized the manuscript, collected data, developed all software, performed all analyses, and wrote the manuscript. S.B. assessed clinical scans. E.H. assessed clinical scans. All authors reviewed the manuscript. M.D.F. conceptualized the manuscript, provided supervision, and reviewed the manuscript. Acknowledgement C.W.H. would like to acknowledge Rachel Bethune Howard. Data Availability All data is available upon request via ADNI (https://adni.loni.usc.edu/). References Harper L, Bouwman F, Burton EJ, et al. Patterns of atrophy in pathologically confirmed dementias: a voxelwise analysis. J Neurol Neurosurg Psychiatry 2017; 88 : 908–16. Verhagen MV, Guit GL, Hafkamp GJ, Kalisvaart K. The impact of MRI combined with visual rating scales on the clinical diagnosis of dementia: a prospective study. Eur Radiol 2016; 26 : 1716–22. Harper L, Barkhof F, Fox NC, Schott JM. Using visual rating to diagnose dementia: a critical evaluation of MRI atrophy scales. J Neurol Neurosurg Psychiatry 2015; 86 : 1225–33. Harper L, Fumagalli GG, Barkhof F, et al. MRI visual rating scales in the diagnosis of dementia: evaluation in 184 post-mortem confirmed cases. Brain 2016; 139 : 1211–25. Fox NC, Schott JM. Imaging cerebral atrophy: normal ageing to Alzheimer’s disease. The Lancet 2004; 363 : 392–4. Virhammar J, Laurell K, Cesarini KG, Larsson E-M. Preoperative Prognostic Value of MRI Findings in 108 Patients with Idiopathic Normal Pressure Hydrocephalus. Am J Neuroradiol 2014; 35 : 2311–8. Rudie JD, Rauschecker AM, Nabavizadeh SA, Mohan S. Neuroimaging of Dilated Perivascular Spaces: From Benign and Pathologic Causes to Mimics. J Neuroimaging 2018; 28 : 139–49. Dale AM, Fischl B, Sereno MI. Cortical Surface-Based Analysis. NeuroImage 1999; 9 : 179–94. Statistical Parametric Mapping. Elsevier, 2007 DOI:10.1016/B978-0-12-372560-8.X5000-1. Ossenkoppele R, Cohn‐Sheehy BI, La Joie R, et al. Atrophy patterns in early clinical stages across distinct phenotypes of A lzheimer’s disease. Hum Brain Mapp 2015; 36 : 4421–37. Tetreault AM, Phan T, Orlando D, et al. Network localization of clinical, cognitive, and neuropsychiatric symptoms in Alzheimer’s disease. Brain 2020; 143 : 1249–60. Hardcastle C, O’Shea A, Kraft JN, et al. Contributions of Hippocampal Volume to Cognition in Healthy Older Adults. Front Aging Neurosci 2020; 12 : 593833. Van Loenhoud AC, Van Der Flier WM, Wink AM, et al. Cognitive reserve and clinical progression in Alzheimer disease: A paradoxical relationship. Neurology 2019; 93 . DOI:10.1212/WNL.0000000000007821. Egger K, Rau A, Yang S, et al. Automated voxel- and region-based analysis of gray matter and cerebrospinal fluid space in primary dementia disorders. Brain Res 2020; 1739 : 146800. Arbabshirani MR, Plis S, Sui J, Calhoun VD. Single subject prediction of brain disorders in neuroimaging: Promises and pitfalls. NeuroImage 2017; 145 : 137–65. Park J-Y, Na HK, Kim S, et al. Robust Identification of Alzheimer’s Disease subtypes based on cortical atrophy patterns. Sci Rep 2017; 7 : 43270. Seif M, Ziegler G, Freund P. Progressive Ventricles Enlargement and Cerebrospinal Fluid Volume Increases as a Marker of Neurodegeneration in Patients with Spinal Cord Injury: A Longitudinal Magnetic Resonance Imaging Study. J Neurotrauma 2018; 35 : 2941–6. Chua SE, Cheung C, Cheung V, et al. Cerebral grey, white matter and csf in never-medicated, first-episode schizophrenia. Schizophr Res 2007; 89 : 12–21. Nyatega CO, Qiang L, Adamu MJ, Kawuwa HB. Gray matter, white matter and cerebrospinal fluid abnormalities in Parkinson’s disease: A voxel-based morphometry study. Front Psychiatry 2022; 13 : 1027907. Van Der Kleij LA, De Bresser J, Hendrikse J, Siero JCW, Petersen ET, De Vis JB. Fast CSF MRI for brain segmentation; Cross-validation by comparison with 3D T1-based brain segmentation methods. PLOS ONE 2018; 13 : e0196119. De Vis JB, Zwanenburg JJ, Van Der Kleij LA, et al. Cerebrospinal fluid volumetric MRI mapping as a simple measurement for evaluating brain atrophy. Eur Radiol 2016; 26 : 1254–62. De Vis JB, Zwanenburg JJ, Van Der Kleij LA, et al. Cerebrospinal fluid volumetric MRI mapping as a simple measurement for evaluating brain atrophy. Eur Radiol 2016; 26 : 1254–62. Petersen RC, Aisen PS, Beckett LA, et al. Alzheimer’s Disease Neuroimaging Initiative (ADNI): Clinical characterization. Neurology 2010; 74 : 201–9. O’Bryant SE, Humphreys JD, Smith GE, et al. Detecting Dementia With the Mini-Mental State Examination in Highly Educated Individuals. Arch Neurol 2008; 65 . DOI:10.1001/archneur.65.7.963. O’Bryant SE. Staging Dementia Using Clinical Dementia Rating Scale Sum of Boxes Scores: A Texas Alzheimer’s Research Consortium Study. Arch Neurol 2008; 65 : 1091. Elwood RW. The Wechsler Memory Scale?Revised: Psychometric characteristics and clinical application. Neuropsychol Rev 1991; 2 : 179–201. American Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition. American Psychiatric Association, 2013 DOI:10.1176/appi.books.9780890425596. Winblad B, Palmer K, Kivipelto M, et al. Mild cognitive impairment - beyond controversies, towards a consensus: report of the International Working Group on Mild Cognitive Impairment. J Intern Med 2004; 256 : 240–6. McKhann G, Drachman D, Folstein M, Katzman R, Price D, Stadlan EM. Clinical diagnosis of Alzheimer’s disease: Report of the NINCDS‐ADRDA Work Group* under the auspices of Department of Health and Human Services Task Force on Alzheimer’s Disease. Neurology 1984; 34 : 939–939. Jack CR, Bernstein MA, Fox NC, et al. The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. J Magn Reson Imaging 2008; 27 : 685–91. Scheltens P, Launer LJ, Barkhof F, Weinstein HC, Gool WA. Visual assessment of medial temporal lobe atrophy on magnetic resonance imaging: Interobserver reliability. J Neurol 1995; 242 : 557–60. Rau A, Urbach H. The MTA score—simple and reliable, the best for now? Eur Radiol 2021; 31 : 9057–9. Claus JJ, Staekenborg SS, Holl DC, et al. Practical use of visual medial temporal lobe atrophy cut-off scores in Alzheimer’s disease: Validation in a large memory clinic population. Eur Radiol 2017; 27 : 3147–55. Pasquier F, Leys D, Weerts JGE, Mounier-Vehier F, Barkhof F, Scheltens P. Inter-and Intraobserver Reproducibility of Cerebral Atrophy Assessment on MRI Scans with Hemispheric Infarcts. Eur Neurol 1996; 36 : 268–72. From the Imaging Cognitive Impairment Network (ICINET), Wahlund L-O, Westman E, et al. Imaging biomarkers of dementia: recommended visual rating scales with teaching cases. Insights Imaging 2017; 8 : 79–90. Gaser C, Dahnke R, Thompson PM, Kurth F, Luders E, Alzheimer’s Disease Neuroimaging Initiative. CAT – A Computational Anatomy Toolbox for the Analysis of Structural MRI Data. Neuroscience, 2022 DOI:10.1101/2022.06.11.495736. Manjón JV, Coupé P, Martí‐Bonmatí L, Collins DL, Robles M. Adaptive non‐local means denoising of MR images with spatially varying noise levels. J Magn Reson Imaging 2010; 31 : 192–203. Ashburner J, Friston KJ. Unified segmentation. NeuroImage 2005; 26 : 839–51. Klein A, Andersson J, Ardekani BA, et al. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage 2009; 46 : 786–802. Ashburner J. A fast diffeomorphic image registration algorithm. NeuroImage 2007; 38 : 95–113. Pereira JMS, Williams GB, Acosta-Cabronero J, et al. Atrophy patterns in histologic vs clinical groupings of frontotemporal lobar degeneration. Neurology 2009; 72 : 1653–60. Caspers J, Heeger A, Turowski B, Rubbert C. Automated age- and sex-specific volumetric estimation of regional brain atrophy: workflow and feasibility. Eur Radiol 2021; 31 : 1043–8. Plant C, Teipel SJ, Oswald A, et al. Automated detection of brain atrophy patterns based on MRI for the prediction of Alzheimer’s disease. NeuroImage 2010; 50 : 162–74. Schultz V, Hedderich DM, Schmitz-Koep B, et al. Removing outliers from the normative database improves regional atrophy detection in single-subject voxel-based morphometry. Neuroradiology 2024; 66 : 507–19. Shiino A, Watanabe T, Maeda K, Kotani E, Akiguchi I, Matsuda M. Four subgroups of Alzheimer’s disease based on patterns of atrophy using VBM and a unique pattern for early onset disease. NeuroImage 2006; 33 : 17–26. Chetelat G, Desgranges B, Landeau B, et al. Direct voxel-based comparison between grey matter hypometabolism and atrophy in Alzheimer’s disease. Brain 2007; 131 : 60–71. La Joie R, Perrotin A, Barré L, et al. Region-Specific Hierarchy between Atrophy, Hypometabolism, and β-Amyloid (Aβ) Load in Alzheimer’s Disease Dementia. J Neurosci 2012; 32 : 16265–73. Perry DC, Brown JA, Possin KL, et al. Clinicopathological correlations in behavioural variant frontotemporal dementia. Brain 2017; 140 : 3329–45. Kueper JK, Speechley M, Montero-Odasso M. The Alzheimer’s Disease Assessment Scale–Cognitive Subscale (ADAS-Cog): Modifications and Responsiveness in Pre-Dementia Populations. A Narrative Review. J Alzheimers Dis 2018; 63 : 423–44. Ma J, Zheng M-X, Wu J-J, et al. Mapping the long-term delayed recall-based cortex-hippocampus network constrained by the structural and functional connectome: a case-control multimodal MRI study. Alzheimers Res Ther 2023; 15 : 61. Foster JK, Meikle A, Goodson G, et al. The Hippocampus and Delayed Recall: Bigger is not Necessarily Better? Memory 1999; 7 : 715–33. Kramer JH, Schuff N, Reed BR, et al. Hippocampal volume and retention in Alzheimer’s disease. J Int Neuropsychol Soc 2004; 10 : 639–43. Ma J, Wu J-J, Xing X-X, et al. Circuit-based neuromodulation enhances delayed recall in amnestic mild cognitive impairment. J Neurol Neurosurg Psychiatry 2024; : jnnp-2023-333152. Mortimer JA, Gosche KM, Riley KP, Markesbery WR, Snowdon DA. Delayed recall, hippocampal volume and Alzheimer neuropathology: Findings from the Nun Study. Neurology 2004; 62 : 428–32. Winkler AM, Ridgway GR, Webster MA, Smith SM, Nichols TE. Permutation inference for the general linear model. NeuroImage 2014; 92 : 381–97. Raschka S. Model Evaluation, Model Selection, and Algorithm Selection in Machine Learning. 2018. DOI:10.48550/ARXIV.1811.12808. Whang Y-J. Consistent bootstrap tests of parametric regression functions. J Econom 2000; 98 : 27–46. Barnett AJ, Guo Z, Jing J, et al. Improving Clinician Performance in Classifying EEG Patterns on the Ictal–Interictal Injury Continuum Using Interpretable Machine Learning. NEJM AI 2024; 1 . DOI:10.1056/AIoa2300331. Seabold S, Perktold J. Statsmodels: Econometric and Statistical Modeling with Python. In: Proceedings of the 9th Python in Science Conference. 2010: 92–6. Abraham A, Pedregosa F, Eickenberg M, et al. Machine learning for neuroimaging with scikit-learn. Front Neuroinformatics 2014; 8 . DOI:10.3389/fninf.2014.00014. Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine Learning in Python. J Mach Learn Res 2011; 12 : 2825–30. Vargha-Khadem F, Gadian DG, Watkins KE, Connelly A, Van Paesschen W, Mishkin M. Differential Effects of Early Hippocampal Pathology on Episodic and Semantic Memory. Science 1997; 277 : 376–80. Yoshii F, Kawaguchi C, Kohara S, et al. Characteristic deterioration of ADAS-Jcog subscale scores and correlations with regional cerebral blood flow reductions in Alzheimer’s disease. Neurol Sci 2018; 39 : 909–18. on behalf of the MAGNIMS study group, Sastre-Garriga J, Pareto D, et al. MAGNIMS consensus recommendations on the use of brain and spinal cord atrophy measures in clinical practice. Nat Rev Neurol 2020; 16 : 171–82. Hobden G, Colbourne E, Pendlebury ST, Demeyere N. Reliability of the global cortical atrophy visual rating scale applied to computed tomography versus magnetic resonance imaging scans in acute stroke. Neurol Sci 2024; 45 : 1549–56. Cardinale F, Chinnici G, Bramerio M, et al. Validation of FreeSurfer-Estimated Brain Cortical Thickness: Comparison with Histologic Measurements. Neuroinformatics 2014; 12 : 535–42. Yao Z, Hu B, Liang C, Zhao L, Jackson M, the Alzheimer’s Disease Neuroimaging Initiative. A Longitudinal Study of Atrophy in Amnestic Mild Cognitive Impairment and Normal Aging Revealed by Cortical Thickness. PLoS ONE 2012; 7 : e48973. Oxenford S, Ríos AS, Hollunder B, et al. WarpDrive: Improving spatial normalization using manual refinements. Med Image Anal 2024; 91 : 103041. Pereira JMS, Xiong L, Acosta-Cabronero J, Pengas G, Williams GB, Nestor PJ. Registration accuracy for VBM studies varies according to region and degenerative disease grouping. NeuroImage 2010; 49 : 2205–15. Additional Declarations No competing interests reported. Supplementary Files howardcsfnpjasupp.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6423862","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":450250965,"identity":"6e9b8759-e184-4f6c-aba5-f62de9d73029","order_by":0,"name":"Calvin W. Howard","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/UlEQVRIie3OMWvCQBTA8RcOmuVB1gtR+hUiBzpom88igcsSROjiWOganE/Mhyh07XBwEBe7R3TOlEKhkyjYM24tZ1snh/sPx+PBj3cANts1RpuXoOdmQADkaYsAzuMvpOVny3+SQVimfyTe/Kmg29c+wuq9CMawue2JuPqoYdB+loYjm4L7WZWgMx/xQEDVyUvOZjlwZiIhTbslSoUkSLtrBOUIlIzoYXiOrPaa3PjLhkQCF5+aHM6S9fEKUmzIULjZ8Yo0ElryZNeSiRajZIehijV5cPIwZjMD8USsOrXsR9HiTTGcqDtB3BeoJ/ftqYF8/+ePwWaz2WyX9AXYTFv2VepmBwAAAABJRU5ErkJggg==","orcid":"","institution":"Harvard Medical School","correspondingAuthor":true,"prefix":"","firstName":"Calvin","middleName":"W.","lastName":"Howard","suffix":""},{"id":450250966,"identity":"2bd31822-a922-4278-9012-fe2e7624a02c","order_by":1,"name":"Sheena Barotono","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Sheena","middleName":"","lastName":"Barotono","suffix":""},{"id":450250967,"identity":"49b3c4ff-8f02-4bc6-b10f-f548f8aade35","order_by":2,"name":"Elmira Hassanzadeh","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Elmira","middleName":"","lastName":"Hassanzadeh","suffix":""},{"id":450250968,"identity":"d2256a1b-2911-45db-b5ce-5700f02e35ab","order_by":3,"name":"Grace Burt","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Grace","middleName":"","lastName":"Burt","suffix":""},{"id":450250969,"identity":"e1407271-91aa-4224-980e-b6a432697f8e","order_by":4,"name":"Joseph Turner","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Joseph","middleName":"","lastName":"Turner","suffix":""},{"id":450250970,"identity":"424ec298-3f69-418a-8d3c-76d7a1cb7d87","order_by":5,"name":"Helen Friedrich","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Helen","middleName":"","lastName":"Friedrich","suffix":""},{"id":450250971,"identity":"bdc0e00c-588b-47e4-97a5-a2aea25e31a6","order_by":6,"name":"Maximillian U. Friedrich","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Maximillian","middleName":"U.","lastName":"Friedrich","suffix":""},{"id":450250972,"identity":"149bec9d-64cc-463e-844d-09cf6b89c877","order_by":7,"name":"Tony Phan","email":"","orcid":"","institution":"Vanderbilt University Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Tony","middleName":"","lastName":"Phan","suffix":""},{"id":450250973,"identity":"57c3405a-84cb-4f65-ab90-15df19064345","order_by":8,"name":"William Drew","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"William","middleName":"","lastName":"Drew","suffix":""},{"id":450250974,"identity":"774feb5e-6e50-41f1-89e9-5e8283339fb7","order_by":9,"name":"Sara Lariviere","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Sara","middleName":"","lastName":"Lariviere","suffix":""},{"id":450250975,"identity":"e2055162-4ce2-4b3e-8d8a-4e970225ac9a","order_by":10,"name":"Thomas Picht","email":"","orcid":"","institution":"Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin","correspondingAuthor":false,"prefix":"","firstName":"Thomas","middleName":"","lastName":"Picht","suffix":""},{"id":450250976,"identity":"261303e9-ed8f-4332-8174-c2458b592a50","order_by":11,"name":"Katharina Faust","email":"","orcid":"","institution":"Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin","correspondingAuthor":false,"prefix":"","firstName":"Katharina","middleName":"","lastName":"Faust","suffix":""},{"id":450250977,"identity":"94599aad-9910-4bf5-91ef-c68221e298d8","order_by":12,"name":"Michael D. Fox","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"D.","lastName":"Fox","suffix":""}],"badges":[],"createdAt":"2025-04-11 01:53:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6423862/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6423862/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82204735,"identity":"e6765b93-dd9d-4bf0-abf2-06076365bafd","added_by":"auto","created_at":"2025-05-07 16:58:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":320792,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eUsing CSF-based atrophy to identify atrophy in individual patients. \u003c/strong\u003eBrain atrophy results in expansion of CSF spaces. Abnormal CSF space enlargement can be detected in individual patients by comparison to a cohort of normative control subjects. The location of the CSF-based atrophy in an individual patient can then be overlaid on an intact reference brain to identify voxels that are classified as CSF in the individual patient but as brain tissue in healthy individuals.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/3cada766c6c2e58ec1465e19.png"},{"id":82203982,"identity":"7a6b39cc-424b-4856-abf2-3c51a1e10363","added_by":"auto","created_at":"2025-05-07 16:50:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":309345,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExample of atrophy in a single patient identified by different methods. \u003c/strong\u003eA single patient with AD and isolated biparietal atrophy based on clinical ratings (left column, arrows) was analyzed using four automated methods for atrophy detection (other columns). Axial views of the patients’ MRI are shown in the top row and a 3D rendering is shown in the bottom row. Automated methods based on cerebrospinal fluid (blue column) aligned better with clinical ratings than other automated methods based on cortical thickness (orange), white matter (green column), or grey matter (red column). Axial images are shown in radiographic convention.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/d4d98c276dbc7eb3a15eee4e.png"},{"id":82203984,"identity":"aa2ba7b7-80e3-4018-b735-30e7a7f0b918","added_by":"auto","created_at":"2025-05-07 16:50:40","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":898953,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCSF correlates most strongly with neuroradiologist evaluations. \u003c/strong\u003eA) Boxplots of how well each automated method’s atrophy correlates with neuroradiologist atrophy ratings. Each point in the boxplot is the value of a single lobe’s correlation between automatic atrophy and neuroradiologist. There was a significant difference of correlation strength across the automated methods (Kruskal-Wallis, p = 0.0087). B) Same data, but with correlations shown by lobe. Bars represent the correlation between automatically detected atrophy and clinical rating. Permutation testing found significant differences across methods within lobes (stars: * = \u003cem\u003ep\u003c/em\u003e \u0026lt; .05; ** = \u003cem\u003ep\u003c/em\u003e \u0026lt; .01; *** = \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). Dashed line is the threshold correlations must achieve to be significant for this sample size.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/ee089407fd2ff109b032141a.png"},{"id":82203987,"identity":"d485c434-6e08-439d-b1f7-8582d898070c","added_by":"auto","created_at":"2025-05-07 16:50:40","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":963162,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCSF relates most strongly to cognitive symptoms. \u003c/strong\u003eA) Boxplots of how well each automated method’s atrophy correlates ADAS-Cog 11. Each point in the boxplot is the value of a single lobe’s correlation between automatic atrophy and neuroradiologist. There was a significant difference of correlation strength across the automated methods (Kruskal-Wallis, p = 0.0087). B) Same data, but with correlations shown by lobe. Bars represent the correlation between automatically detected atrophy and ADAS-Cog 11. Permutation testing found significant differences across methods within lobes (stars: * = \u003cem\u003ep\u003c/em\u003e \u0026lt; .05; ** = \u003cem\u003ep\u003c/em\u003e \u0026lt; .01; *** = \u003cem\u003ep\u003c/em\u003e \u0026lt; .001). Dashed line is the threshold correlations must achieve to be significant for this sample size.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/13ae1049beffbbf72a4cd938.png"},{"id":82203986,"identity":"b7b13852-7136-4361-8faf-cef9217cad6b","added_by":"auto","created_at":"2025-05-07 16:50:40","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":289736,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCSF best localizes verbal memory function to the left hippocampus.\u003c/strong\u003e A) Whole-brain peak correlations between atrophy and delayed word recall. CSF was the only method to identify a whole-brain peak within the expected neuroanatomical location of the left hippocampus (left column) and the only method to identify voxel-wise correlations that survived family-wise error correction for multiple comparisons (outlined in red).\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/721e35d180319097e383da50.png"},{"id":82203989,"identity":"3e180f90-759a-4568-a5b6-66fbd1c05752","added_by":"auto","created_at":"2025-05-07 16:50:40","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":295566,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCSF best distinguishes between Alzheimer, MCI, and control patients. \u003c/strong\u003eA)\u003cstrong\u003e \u003c/strong\u003eValidation test of logistic regression on 150 unseen patients. B) Bootstrapped 95% confidence intervals of the validation dataset for each model. The CSF AUROC was significant higher than GM (p = 0.0081), WM (p = 0.038), and CTh (p = 0.0046), by DeLong’s test. Comparison of AUCs at each bootstrap showed CSF outperformed GM 99.21% of the time, WM 96.22% of the time, and CTh 99.54% of the time. Dashed grey line represents chance-level performance. Stars: * = \u003cem\u003ep\u003c/em\u003e \u0026lt; .05; ** = \u003cem\u003ep\u003c/em\u003e \u0026lt; .01; *** = \u003cem\u003ep\u003c/em\u003e \u0026lt; .001\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/6213dd452305a2cf4d8a2c5e.png"},{"id":83618751,"identity":"2709d3a9-8ead-4e40-8ea8-58f1bef7df9a","added_by":"auto","created_at":"2025-05-29 14:32:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4362678,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/edd3bcc1-ea8e-4d6f-b56e-eadef71cf370.pdf"},{"id":82204005,"identity":"b201e70a-eb7c-4753-9cfc-ce90e6643328","added_by":"auto","created_at":"2025-05-07 16:50:41","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":11019422,"visible":true,"origin":"","legend":"","description":"","filename":"howardcsfnpjasupp.docx","url":"https://assets-eu.researchsquare.com/files/rs-6423862/v1/126740b6ba98b977e1496ea4.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Using Cerebrospinal Fluid Improves Detection of Individual Brain Atrophy","fulltext":[{"header":"Introduction","content":"\u003cp\u003eWhen neurologists and neuroradiologists are working up a dementia diagnosis, they routinely assess patient brain scans for disease-specific atrophy patterns.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e In clinical practice, these atrophy patterns are often identified qualitatively,\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e by visually inspecting the contrast between brain and cerebrospinal fluid (CSF) for the hallmark widening of CSF that occurs with brain atrophy.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e Despite this qualitative nature, clinical inspection of CSF to proxy brain atrophy both accurately identifies atrophy location\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e and neurodegenerative diagnosis.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eAlthough it may be useful clinically, using widened CSF spaces to detect atrophy is only a proxy of brain atrophy, not a direct measure of neuronal loss. These widened CSF spaces could represent multiple processes, such as atrophy, increased CSF pressure, or increased glymphatic space.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e As such, most atrophy detection software and research has focused on more direct measures of grey matter (GM) volume, white matter (WM) volume, cortical thickness (CTh).\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e Despite evaluating the brain differently than clinicians, these methods robustly map brain atrophy,\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e explain symptom variance,\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e and aid diagnosis.\u003csup\u003e\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eGiven the established strengths of GM, WM, and CTh, only a handful of research studies have explored atrophy detection software using widened CSF spaces as a proxy for atrophy.\u003csup\u003e\u003cspan additionalcitationids=\"CR18 CR19 CR20\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e In these studies, CSF performed surprisingly well in atrophy detection,\u003csup\u003e\u003cspan additionalcitationids=\"CR18 CR19 CR20\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e disease diagnosis,\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e and sometimes outperformed assessments based on GM, WM, and CTh.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e While these findings are interesting, most of these results required specialized MRI sequences,\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e were restricted to specific brain regions,\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e or focused on group-level findings rather than individual patients.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e Thus, the utility of CSF for assessment of a patient\u0026rsquo;s atrophy remains unclear.\u003c/p\u003e \u003cp\u003eHere, we evaluate an automated approach to map brain atrophy in individual patients based on expansion of CSF spaces visible on routine clinical MRI. By comparing a patient\u0026rsquo;s CSF to a reference \u0026lsquo;healthy\u0026rsquo; distribution, we generate a patient-specific map of CSF-based atrophy as a proxy of the patient\u0026rsquo;s underlying atrophy. We then tested the clinical relevance of this map using four outcome metrics: 1) correlation with clinician evaluations, 2) correlation with clinical symptoms, 3) localization potential, and 4) diagnostic potential. For all outcome metrics, we compared results to conventional atrophy detection algorithms based on GM, WM, and CTh.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eEthics Statement\u003c/h2\u003e \u003cp\u003e This study was conducted in accordance with ethical standards and approved by the Institutional Review Board of the Brigham and Women\u0026rsquo;s Hospital and Harvard Medical School, Boston, Massachusetts. Given the secondary use of research data, the study was exempted from obtaining informed consent.\u003c/p\u003e \u003cp\u003eData used in this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu).\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e Institutional Review Board (IRB) approval was obtained from all sites, and informed consent obtained from all participants. Informed written consent was collected from all subjects within each site. ADNI provided access to all data used in this study and the authors used their guidelines for acknowledging them in the author section.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eSubjects\u003c/h3\u003e\n\u003cp\u003eAll subjects were recruited from ADNI1 (2004\u0026ndash;2009).\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e Each patient received an initial visit with a physical examination, neurological examination, neuropsychological examination, fluid biomarkers, and neuroimaging.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e At the end of this visit, they received a diagnosis of cognitively normal (control), mild cognitive impairment (MCI), or Alzheimer Disease (AD).\u003c/p\u003e \u003cp\u003eDiagnosis was based on clinical criteria demonstrating a combination of Mini Mental Status Examination Score\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, Clinical Dementia Rating score\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, Wechsler Memory Scale-Revised\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, with additional criteria for: dementia\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e, MCI\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e, and AD\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e (Supplementary Table\u0026nbsp;1). A detailed breakdown of inclusion/exclusion criteria used in ADNI\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e is available (Supplementary Table\u0026nbsp;2).\u003c/p\u003e \u003cp\u003eOverall, 338 subjects were included from ADNI (74.0\u0026thinsp;\u0026plusmn;\u0026thinsp;6.8 years, 48.5% female). To compare patient brains against a reference distribution, we collected a reference cohort of \u0026lsquo;healthy\u0026rsquo; controls (n\u0026thinsp;=\u0026thinsp;138, 73.7\u0026thinsp;\u0026plusmn;\u0026thinsp;6.4 years, 48.5% female). We also collected a separate cohort of patients for evaluation (n\u0026thinsp;=\u0026thinsp;200, 73.0\u0026thinsp;\u0026plusmn;\u0026thinsp;7.0 years, 47.5% female), composed of controls (n\u0026thinsp;=\u0026thinsp;68, 72.4\u0026thinsp;\u0026plusmn;\u0026thinsp;7.1 years, 42.9% female), MCI patients (n\u0026thinsp;=\u0026thinsp;72, 72.4\u0026thinsp;\u0026plusmn;\u0026thinsp;7.1 years, 42.9% female), and AD patients (n\u0026thinsp;=\u0026thinsp;60, 74.5\u0026thinsp;\u0026plusmn;\u0026thinsp;6.9 years, 54% female). The patient cohorts were randomly subsampled from ADNI in a 1:1:1 ratio.\u003c/p\u003e \u003cp\u003eWe then split the 200 patients into discovery (n\u0026thinsp;=\u0026thinsp;50, for initially assessing the algorithm\u0026rsquo;s utility) and validation cohorts (n\u0026thinsp;=\u0026thinsp;150, for testing the algorithm in unseen patients). The discovery cohort (n\u0026thinsp;=\u0026thinsp;50, 73.9\u0026thinsp;\u0026plusmn;\u0026thinsp;7.8 years, 36.7% female) was composed of control (n\u0026thinsp;=\u0026thinsp;21), MCI (n\u0026thinsp;=\u0026thinsp;20), and AD patients (n\u0026thinsp;=\u0026thinsp;9). The validation cohort (n\u0026thinsp;=\u0026thinsp;150, 74.8\u0026thinsp;\u0026plusmn;\u0026thinsp;6.9 years, 51.3% female) was composed of control (n\u0026thinsp;=\u0026thinsp;47), MCI (n\u0026thinsp;=\u0026thinsp;52), and AD patients (n\u0026thinsp;=\u0026thinsp;51). Demographic characteristics of this cohort are available (Supplementary Table\u0026nbsp;3).\u003c/p\u003e\n\u003ch3\u003eNeuroimaging\u003c/h3\u003e\n\u003cp\u003eThe MRI protocol for ADNI1 (2004\u0026ndash;2009) acquired structural imaging on 3T scanners using T1 sequences (MPRAGE protocol: sagittal plane, TR/TE/TI, 2400/3/1000 ms, flip angle 8\u0026deg;, 24 cm FOV, 192 \u0026times; 192 in-plane matrix, 1.2 mm slice thickness).\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e In this study, we evaluated only T1 scans acquired on 3T scanners.\u003c/p\u003e\n\u003ch3\u003eClinician Evaluation of Atrophy\u003c/h3\u003e\n\u003cp\u003eA neuroradiologist (EH), a cognitive neurologist (SB), and a clinical neuroimaging expert (CH) assessed 2100 brain regions across 200 patient brain scans, spanning 1050 regions in the 50 discovery cohort patients (EH, SB, CH assessed 7 regions across 50 patients), and 1050 regions in the 150 validation cohort patients (CH assessed 7 regions across 150 patients). All clinicians were blinded, rated independently of each other, and instructed in use of the appropriate visual rating scale. EH, SB, and CH evaluated the validation cohort. CH evaluated the validation cohort. All clinician evaluations of atrophy occurred before quantitative atrophy measurements. No clinical information was provided (i.e. age), ensuring raters evaluated the brain scans in isolation.\u003c/p\u003e\n\u003ch3\u003eVisual Rating Scales\u003c/h3\u003e\n\u003cp\u003eIn all patients, seven regions were evaluated: the frontal lobe, temporal lobe, parietal lobe, occipital lobe, mesial temporal lobe, and cerebellum; ventriculomegaly was also graded. In cases of asymmetric atrophy, the worse grade was taken.\u003c/p\u003e \u003cp\u003eThe mesial temporal lobe was graded using the Mesial Temporal Atrophy (MTA) scale,\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e the most reliable and widely used visual scale for mesial temporal atrophy.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e,\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e The MTA scale ranges from 0 (intact) to 4 (severe). The full MTA scale is available in the supplements (Supplementary Table\u0026nbsp;4).\u003c/p\u003e \u003cp\u003eThe other 6 out of 7 regions were evaluated with the Global Cortical Atrophy (GCA) scale,\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e as it is the most widely used assessment scale of brain-wide atrophy.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e was applied to all other regions. Each region was evaluated independently of other regions, as has been previously done.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e The GCA scale ranges from 1 (intact) to 4 (severe). The full GCA scale is available in the supplements (Supplementary Table\u0026nbsp;5).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eReliability of Clinician Evaluations\u003c/h2\u003e \u003cp\u003eThe interrater reliability among the 3 clinicians was evaluated across all regions using intraclass correlation coefficient (ICC) and comparison of atrophy rating distributions via Kolmogorov Smirnov Test. Author CH was compared specifically to the neuroradiologist (author EH) to ensure validity of ratings of the validation cohort.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eVoxel Based Morphometry\u003c/h3\u003e\n\u003cp\u003eVBM analyses were performed with Statistical Parametric Mapping (SPM) and Computational Anatomy Toolbox (CAT12) to derive maps of each patient\u0026rsquo;s GM, WM, and CSF segments as previously described.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e Briefly, all 3D T1-weighted MRI scans are preprocessed with spatial adaptive non-local means (SANLM) denoising filter.\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e These are then bias corrected and affine registered to the space of SPM\u0026rsquo;s tissue probability maps.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e Next, GM, WM, and CSF components are segmented out using the SPM tissue probability maps,\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e resulting in probability maps for each segment.\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e Diffeomorphic Anatomic Registration Through Exponentiated Lie Algebra (DARTEL) is used to normalize the segmented scans into MNI space.\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e,\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e DARTEL provides precise normalization to MNI space, and has the advantage of preserving structural anomalies inherent to disease processes.\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e,\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e This generates a map of \u0026lsquo;concentration\u0026rsquo; of each segment of the brain across all voxels in MNI space. The Jacobian determinant derived from the warp is then applied to the voxels, modulating them and converting them to an adjusted volume which accounts for the distortion by the warp process.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e The resulting tissue segment maps were smoothed with both 0mm and 6mm isotropic smoothing kernels. This process results in the final grey, white, and cerebrospinal fluid volume maps.\u003c/p\u003e\n\u003ch3\u003eSurface Based Morphometry\u003c/h3\u003e\n\u003cp\u003eSBM analyses were performed with FreeSurfer 5.3.0 to derive CTh values across each patient\u0026rsquo;s brain.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e In brief, each patient\u0026rsquo;s brain was affine-registered to MNI space. Subsequently, WM segmentation is performed to delineate the edge of the WM, which is then tessellated into a surface. The outer boundary of the cortex is then estimated by identifying the CSF edge. The radial distance between the WM surface and the outer boundary of the cortex is then estimated at all points, representing the thickness of the cortex (CTh).\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eMapping Patient Atrophy\u003c/h2\u003e \u003cp\u003eThere are two established methods for defining atrophy in reference to a control distribution: Z-scoring\u003csup\u003e\u003cspan additionalcitationids=\"CR43 CR44 CR45\" citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e and W-scoring\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e,\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Briefly, we calculate voxelwise Z- and W-scores by defining a normal distribution based on the 138 healthy control subjects, then evaluating the voxelwise deviation in experimental subjects. Z-scoring simply compares each patient\u0026rsquo;s brain to the expected mean, and normalizes the difference based on standard deviation (Supplementary Eq.\u0026nbsp;1). W-scoring fits a linear model to the control distribution, then uses that to predict the expected value of the patient and normalizes this by the residual standard deviation (Supplementary Eq.\u0026nbsp;2). We derived both for all patients, but present Z-scores in the primary results as they were ultimately found to be superior across all evaluation metrics.\u003c/p\u003e \u003cp\u003eThe Z- or W-maps can be thresholded at a value of 2, representing volume loss beyond the 98th percentile of expected. This is our operational definition of atrophy. GM, WM, and CT are thresholded below a value of -2 to identify atrophy, while CSF is thresholded above +\u0026thinsp;2. Composite atrophy maps were created by summating thresholded atrophy maps. All combinations of composite maps across each method (CSF, GM, WM, CTh) were created. We generated atrophy maps for all patients across smoothing kernels of 0mm and 6mm. Reference controls were smoothed to 6mm prior to comparing patients against them.\u003c/p\u003e \u003cp\u003eTo map expansion of CSF spaces, we compared the CSF in each patient to a normative distribution of CSF from healthy controls (Supplementary Fig.\u0026nbsp;8). The result is a voxel-wise map, where values represent deviations in the patient relative to the normal distribution (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Note that this map of CSF-based atrophy is only an indirect proxy of underlying brain atrophy, we will refer to it as a \u0026ldquo;CSF-based atrophy map\u0026rdquo; to match the terminology used for other tissue components \u0026ldquo;e.g. GM-based atrophy map\u0026rdquo;.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eMeasuring Degree of Atrophy in Each Brain Region\u003c/h2\u003e \u003cp\u003eAfter deriving either Z- or W-maps of each patient\u0026rsquo;s atrophy, we evaluated atrophy in all 7 regions of interest (mesial temporal, frontal, temporal, parietal, and occipital lobes, as well as cerebellum, and periventricular subcortex). Z- or W-scores within each region were summated. This provided an evaluation of quantitative atrophy in each region of interest.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eRelating Atrophy to Clinician Assessments\u003c/h2\u003e \u003cp\u003eGiven the ordinal nature of the visual rating scales, Spearman Correlation related quantified atrophy to clinician atrophy scores. For every lobe, across each method, we correlated atrophy to clinician atrophy scores. Kruskal-Wallis tests compared performance of each method, using performance across the 7 ROIs to estimate the distribution of the method\u0026rsquo;s overall performance. Post-hoc testing used permutation analysis to compare significant difference between correlations across methods. ANOVA and permutation testing evaluated differences within each region, across methods. Comparisons were made with and without the ventricles and subcortex regions of interest to ensure fair comparison to CTh. All evaluations presented in the figures are relative to the clinical neuroradiologist and robust when compared to the neurologist or resident.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eRelating Atrophy to Cognitive Outcomes\u003c/h2\u003e \u003cp\u003eTo test whether atrophy was related to clinical symptoms, we examined the relationship between regional brain atrophy and each patient\u0026rsquo;s baseline Alzheimer\u0026rsquo;s Disease Assessment Scale of Cognition-11 (ADAS-Cog 11). The ADAS-Cog 11 was chosen as it is the most comprehensive metric of cognitive function in Alzheimer\u0026rsquo;s, and is ubiquitously available in ADNI.\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e Given the ordinal nature of the ADAS-Cog 11, Spearman correlations related atrophy to each patient\u0026rsquo;s ADAS-Cog 11. Kruskal-Wallis tests compared performance of each method, using performance across the 7 ROIs to estimate the distribution of the method\u0026rsquo;s overall performance. Post-hoc testing used permutation analysis to compare significant difference between correlations across methods. ANOVA and permutation testing evaluated differences within each region, across methods. Comparisons were made with and without the ventricles and subcortex regions of interest to ensure fair comparison to CTh.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eUsing Atrophy to Localize Delayed Word Recall\u003c/h2\u003e \u003cp\u003eWe assessed if the atrophy methods could localize a neurological function that was both highly focal and known to be affected in neurodegeneration. We chose delayed word recall (ADAS-Cog 11 question 4), a function of memory impaired by Alzheimer\u0026rsquo;s, which is known to localize primarily to the left hippocampus by histopathological study\u003csup\u003e\u003cspan additionalcitationids=\"CR51 CR52 CR53\" citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e The ability for each method to localize delayed word recall was measured via whole-brain data-driven analysis and in an a priori region of interest analysis.\u003c/p\u003e \u003cp\u003eThe whole-brain analysis used voxel- or vertex-wise Spearman Correlations to relate atrophy at each point to a patient\u0026rsquo;s verbal recall. Permutation testing shuffled outcomes to estimate p-value, which were maximum statistic family-wise error corrected.\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e The difference in peak hippocampal correlations were also compared across methods using permutation analysis.\u003c/p\u003e \u003cp\u003eA hippocampal region of interest was used to quantify atrophy for each patient. Atrophy within this hippocampal region was then correlated with each patient\u0026rsquo;s delayed word recall. Permutation testing compared the strength of this correlation across methods.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eUsing Atrophy to Diagnose Patients as Control, MCI, or AD\u003c/h2\u003e \u003cp\u003eWe also assessed how well each method could detect diagnostically relevant atrophy. To do this, we employed a standard method of discriminating between patients: by calculating atrophy within regions of interest and using it to fit a logistic regression on diagnosis.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e Atrophy within the frontal, parietal, temporal, occipital, and mesial temporal lobes were quantified for each disease. Then, a multiclass logistic regression related atrophy in these regions to classify patients as control, MCI, or AD.\u003c/p\u003e \u003cp\u003e The logistic regressions were fit on our well-characterized discovery dataset that all clinicians agreed upon (n\u0026thinsp;=\u0026thinsp;50) and were subsequently tested on the held-out validation dataset (n\u0026thinsp;=\u0026thinsp;150). Performance was measured by each model\u0026rsquo;s area under the receiver operating characteristic (AUROC). Bootstrapping (n\u0026thinsp;=\u0026thinsp;10 000) was used to estimate confidence intervals for each AUROC. A smaller training dataset than validation dataset is expected to be more rigorous, and is known to produce lower AUROC estimates.\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e Significant differences between bootstrapped confidence intervals were evaluated using two methods: 1) DeJong\u0026rsquo;s test,\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e and 2) proportion of superior samples.\u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eStatistics Statement\u003c/h2\u003e \u003cp\u003eStatistical analyses were conducted in Python 3.10 using Statsmodels 0.14.0, Nilearn 0.10.1, and Scikit-learn 1.3.0.\u003csup\u003e\u003cspan additionalcitationids=\"CR60\" citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e Descriptive statistics are presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation. Correlations used Spearman's methods as appropriate for ordinal data. Logistic regressions were performed with multiclass multinomial logistic regressions. Comparison of atrophy distributions was performed with Kolmogorov-Smirnov tests with Bonferroni family wise error correction.\u003c/p\u003e \u003cp\u003ePermutation testing was performed as previously described.\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e Briefly, a statistic of interest (Spearman R) was calculated on the actual data. Then, the outcomes were shuffled and the statistic was recomputed; this occurred multiple times (n\u0026thinsp;=\u0026thinsp;10 000). The number of times the random process exceeded the observed value, averaged, provides the p-value.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eClinician Evaluation of Patient Atrophy\u003c/h2\u003e \u003cp\u003eThree clinical raters scored atrophy across 7 brain regions using standard visual rating scales for 50 subjects in our discovery cohort (n\u0026thinsp;=\u0026thinsp;1050 atrophy scores). Atrophy ratings between a board-certified neuroradiologist and two independent clinical experts were reliable (Supplementary Fig.\u0026nbsp;1) across all brain regions (mean ICC\u0026thinsp;=\u0026thinsp;0.81, max p\u003csub\u003eFWE\u003c/sub\u003e = 0.0036). One clinical expert analyzed the additional 150 validation subjects (n\u0026thinsp;=\u0026thinsp;1050 atrophy scores). This expert\u0026rsquo;s evaluations were validated against the neuroradiologist using ICC (Supplementary Fig.\u0026nbsp;2) and reliably rated atrophy compared to the neuroradiologist (mean ICC\u0026thinsp;=\u0026thinsp;0.82, max p\u003csub\u003eFWE\u003c/sub\u003e = 0.0005).\u003c/p\u003e \u003cp\u003eWe next evaluated if the experts identified the standard distribution of atrophy in AD and MCI patients. AD patients had more atrophy than controls in the frontal, temporal, parietal, and mesial temporal lobes (min U\u0026thinsp;=\u0026thinsp;11, max p\u003csub\u003eFWE\u003c/sub\u003e = 0.0063). MCI patients had more atrophy than controls in the frontal and temporal lobes (min U\u0026thinsp;=\u0026thinsp;8.2, max p\u003csub\u003eFWE\u003c/sub\u003e = 0.041). Rating distributions were not found to be significantly different across raters (p\u003csub\u003eMin\u003c/sub\u003e \u0026gt; 0.05, Supplementary Fig.\u0026nbsp;3). The distribution of atrophy in the validation cohort was not significantly different from the discovery cohort, regardless of which rater was used for comparison (min p\u0026thinsp;=\u0026thinsp;0.33).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eAutomated Evaluation of Individual Atrophy\u003c/h2\u003e \u003cp\u003eWe next generated maps of whole-brain atrophy for each patient based on GM, WM, CTh, and CSF. The atrophy maps often failed to align both with each other and with the clinical ratings, although CSF did align with clinical evaluations (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). GM, WM, and CTh-based methods also detected locations of brain hypertrophy rather than atrophy, although this hypertrophy was not detected by the clinicians nor CSF (Supplementary Fig.\u0026nbsp;4). We next averaged all atrophy maps across 50 AD patients and counted the number of hypertrophic voxels identified. Less than 1% of voxels evaluated with CSF were found to be hypertrophic in these 50 AD patients, compared to 37% in GM, 22% in WM, and 5% in CTh (Supplementary Fig.\u0026nbsp;5).\u003c/p\u003e \u003cp\u003eWe next investigated if processing errors were contributing to the lower performance in GM, WM, and CTh. As a control, we investigated if these segments could recompose the distribution of AD-related atrophy they have previously defined at the group level. We summed and averaged individual atrophy across 50 AD patients, derived by GM, WM, and CTh, and found that these did recomposed the expected precuneus and mesial temporal dominant atrophy with associated frontal and lateral temporal (Supplementary Fig.\u0026nbsp;4).\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e,\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWe next investigated what might be causing higher performance with CSF. At the level of the voxel, CSF improved sensitivity to outlier voxels, which we show in relation to our example patient\u0026rsquo;s whole-brain peak atrophy (Supplementary Fig.\u0026nbsp;7). At the level of the brain, CSF improved whole-brain coverage and enabled detection of missing brain regardless of whether the source of atrophy was from GM, WM, or cortex (Supplementary Fig.\u0026nbsp;8).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eAtrophy Detected by CSF Correlates best with Clinician Ratings\u003c/h2\u003e \u003cp\u003eWe correlated the detected atrophy in each brain region to the atrophy rating from the neuroradiologist (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). Across methods, there was a difference in strength of correlation to the neuroradiologist (Kruskal-Wallis, p\u0026thinsp;=\u0026thinsp;0.0087). Post-hoc permutation testing demonstrated CSF showed the best correlation with clinical ratings (Rho\u003csub\u003emedian\u003c/sub\u003e = 0.50, IQR 0.31\u0026ndash;0.79) and was more correlated to clinician ratings than GM (DRho\u0026thinsp;=\u0026thinsp;0.42, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), WM (DRho\u0026thinsp;=\u0026thinsp;0.41, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), and CTh (DRho\u0026thinsp;=\u0026thinsp;0.30, p\u0026thinsp;=\u0026thinsp;0.0005).\u003c/p\u003e \u003cp\u003eWe repeated the above analyses but assessed each of the 7 brain regions separately (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Correlations to clinician-detected atrophy again varied significantly across CSF, GM, WM, and CTh (ANOVA, p\u003csub\u003eFWE Max\u003c/sub\u003e = 0.0091). Atrophy was significantly correlated with the neuroradiologist rating in for 7/7 regions with CSF (p\u003csub\u003emax\u003c/sub\u003e = 0.043), 0/7 regions with GM (p\u003csub\u003emax\u003c/sub\u003e = 0.72), 0/7 regions with WM (p\u003csub\u003emax\u003c/sub\u003e = 0.91), and 2/7 regions with CTh (p\u003csub\u003emax\u003c/sub\u003e = 0.87). Results were not driven by specific diagnostic subgroups in the discovery cohort (Supplementary Fig.\u0026nbsp;9), controls (Supplementary Fig.\u0026nbsp;10), and were robust across different raters. The analysis was repeated in the full set of 150 test patients, 50 control patients, 50 MCI patients, and 50 AD patients with consistent results (Supplementary Fig.\u0026nbsp;11).\u003c/p\u003e \u003cp\u003eWe next compared different atrophy processing pipelines to understand how CSF performs across them. We found CSF outperformed other methods across various smoothing kernels, but using W-scoring instead of Z-scoring to define atrophy eliminated the performance of CSF (Supplementary Fig.\u0026nbsp;12) and reduced the performance of each atrophy method to below-chance level (Supplementary Fig.\u0026nbsp;13). We tested the differences between Z- and W-scores, finding the difference was not due to controlling for covariates (Supplementary Fig.\u0026nbsp;14), isolating the difference as the model-free nature of Z-scores versus the linear model-based nature of W-scores.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eAtrophy Detected by CSF Correlates with Cognitive Symptoms\u003c/h2\u003e \u003cp\u003eWe next tested whether atrophy detected by each method was correlated with ADAS-Cog 11 scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). There was a significant difference in this correlation across methods (Kruskal-Wallis, p\u0026thinsp;=\u0026thinsp;0.0042). CSF-based atrophy was most correlated with cognitive scores (Rho\u003csub\u003emedian\u003c/sub\u003e = 0.37, IQR 0.34\u0026ndash;0.46) and more correlated to cognitive scores than GM atrophy (DR\u0026thinsp;=\u0026thinsp;0.36, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), WM atrophy (DR\u0026thinsp;=\u0026thinsp;0.30, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), and CTh atrophy (DR\u0026thinsp;=\u0026thinsp;0.22, p\u0026thinsp;=\u0026thinsp;0.0066). Results were robust to using the Clinical Dementia Rating scale sum of boxes (Kruskal-Wallis, p\u0026thinsp;=\u0026thinsp;0.019), Mini Mental Status Exam (Kruskal-Wallis, p\u0026thinsp;=\u0026thinsp;0.0092,) with or without inclusion of the cerebellum and subcortex, and were replicated in the validation cohort (Kruskal-Wallis, p\u0026thinsp;=\u0026thinsp;0.0018).\u003c/p\u003e \u003cp\u003eWe next repeated the correlation of atrophy to symptoms but assessed each of the 7 brain regions independently (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB). The strength of the relationship to clinical symptoms were found to vary significantly across CSF, GM, WM, and CTh (ANOVA, p\u003csub\u003eFWE\u003c/sub\u003e = 0.0029). Atrophy was significantly correlated with clinical symptoms across 7/7 regions using CSF (p\u003csub\u003emax\u003c/sub\u003e = 0.043), 1/7 regions using GM (p\u003csub\u003emax\u003c/sub\u003e = 0.60), 1/7 regions using WM (p\u003csub\u003emax\u003c/sub\u003e = 0.93), and 1/7 regions using CTh (p\u003csub\u003emax\u003c/sub\u003e = 0.18). Control analyses were robust to evaluation in the 150 validation patients, comparing correlation strengths nonparametrically (permutation test, p\u003csub\u003eFWE Max\u003c/sub\u003e = 0.017), and were again optimal with Z-scoring (Supplementary Fig.\u0026nbsp;15) compared to W-scoring (Supplementary Fig.\u0026nbsp;16).\u003c/p\u003e \u003cp\u003eIn additional analyses, we repeated the above analyses but compared each method to the neuroradiologist. Overall, CSF was found to relate more strongly to clinical symptoms than the neuroradiologist (p\u0026thinsp;=\u0026thinsp;0.012, Supplementary Fig.\u0026nbsp;17). When assessing each lobe independently, CSF-based atrophy was found to relate more strongly to symptoms than the neuroradiologist\u0026rsquo;s visual assessment of atrophy in 5/7 lobes (p\u003csub\u003eMax\u003c/sub\u003e = 0.047), which was unique among the automated methods (Supplementary Fig.\u0026nbsp;18).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003eAtrophy Detected by CSF can Localizes specific Neurological deficits\u003c/h2\u003e \u003cp\u003eNext, we wondered if CSF-based atrophy was spatially accurate enough to localize specific neurological deficits. We attempt to localize atrophy impairing delayed word recall, which is known to localize to the left hippocampus.\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e,\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWe used a whole-brain data-driven correlation to relate voxelwise atrophy to delayed word recall performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). This was repeated for each method. CSF was the only method to successfully localize the whole-brain peak (maximum) correlation to the left hippocampus (Rho\u0026thinsp;=\u0026thinsp;0.57, p\u003csub\u003eFWE\u003c/sub\u003e = 0.00302). Permutation testing revealed the CSF peak was significantly stronger than the hippocampal peaks localized by GM (p\u003csub\u003eFWE\u003c/sub\u003e = 0.042), WM (p\u003csub\u003eFWE\u003c/sub\u003e = 0.006), and CTh (p\u003csub\u003eFWE\u003c/sub\u003e = 0.004). This hippocampal localization was specific to delayed word recall and was not similarly localized when this analysis was repeated using the other ten ADAS-Cog 11 subscores. CSF was also the only method to successfully localized delayed word recall after controlling for other cognitive subscores (Supplementary Fig.\u0026nbsp;21) or controlling for age- and sex-related atrophy (Supplementary Fig.\u0026nbsp;22).\u003c/p\u003e \u003cp\u003eWe verified the above results using an a priori hippocampus region of interest. For each method, we quantified left hippocampal atrophy and correlated it to delayed word recall (Supplementary Fig.\u0026nbsp;19). The strongest correlation was seen with CSF atrophy (Rho\u0026thinsp;=\u0026thinsp;0.59, p\u0026thinsp;=\u0026thinsp;0.0000), followed by CTh (Rho\u0026thinsp;=\u0026thinsp;0.47, p\u0026thinsp;=\u0026thinsp;0.0010), GM (Rho\u0026thinsp;=\u0026thinsp;0.31, p\u0026thinsp;=\u0026thinsp;0.039), and WM (Rho\u0026thinsp;=\u0026thinsp;0.24, p\u0026thinsp;=\u0026thinsp;0.11). The correlation between CSF atrophy and verbal memory was significantly stronger than GM (DR\u0026thinsp;=\u0026thinsp;0.28, p\u003csub\u003eFWE\u003c/sub\u003e = 0.013), WM (DR\u0026thinsp;=\u0026thinsp;0.35, p\u003csub\u003eFWE\u003c/sub\u003e = 0.0083), and CTh (DR\u0026thinsp;=\u0026thinsp;0.12, p\u003csub\u003eFWE\u003c/sub\u003e = 0.048). This analysis was replicated in the validation cohort, with CSF again showed the strongest correlation with delayed word recall (Rho\u0026thinsp;=\u0026thinsp;0.68, p\u0026thinsp;=\u0026thinsp;0.0000), followed by CTh (Rho\u0026thinsp;=\u0026thinsp;0.42, p\u0026thinsp;=\u0026thinsp;0.0021), GM (Rho\u0026thinsp;=\u0026thinsp;0.38, p\u0026thinsp;=\u0026thinsp;0.027), and WM (Rho\u0026thinsp;=\u0026thinsp;0.19, p\u0026thinsp;=\u0026thinsp;0.18). CSF was best able to localize additional ADAS-Cog 11 subscores to a-priori lobar localizations (Supplementary Table\u0026nbsp;6).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003eAtrophy Detected by CSF Discriminates Between Diagnoses\u003c/h2\u003e \u003cp\u003eNext, we used the atrophy detected across the frontal, parietal, occipital, temporal, and mesial temporal lobes to distinguish between diseases (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). A logistic regression was trained on the discovery cohort evaluated by the neuroradiologist (n\u0026thinsp;=\u0026thinsp;50), then tested in the held-out validation cohort (n\u0026thinsp;=\u0026thinsp;150). The logistic regression using CSF achieved the highest overall discrimination between control, MCI, and AD patients (AUROC\u0026thinsp;=\u0026thinsp;0.67, p\u0026thinsp;=\u0026thinsp;0.0001), which permutation testing demonstrated was significantly stronger than discrimination from GM (p\u0026thinsp;=\u0026thinsp;0.0081), WM (p\u0026thinsp;=\u0026thinsp;0.038), and CTh (p\u0026thinsp;=\u0026thinsp;0.0046). Differences were robust to DeLong\u0026rsquo;s test (max p\u003csub\u003eFWE\u003c/sub\u003e = 0.001). CSF was found to be more accurate than the radiologist (p\u0026thinsp;=\u0026thinsp;0.012), which was unique among the automated methods (Supplementary Section 6.1).\u003c/p\u003e \u003cp\u003eWe next wondered if the performance of the CSF-based classifier was driven by any specific diagnosis (AD, MCI, or control). We found CSF improved discrimination of AD from MCI patients (AUROC\u0026thinsp;=\u0026thinsp;0.75, p\u0026thinsp;=\u0026thinsp;0.0001), which was significantly improved compared to GM (p\u0026thinsp;=\u0026thinsp;0.0081), WM (p\u0026thinsp;=\u0026thinsp;0.021), and CTh (p\u0026thinsp;=\u0026thinsp;0.012). All methods performed well at classifying AD vs Control, as well as MCI vs Control patients, with no differences between the atrophy methods (Supplementary Fig.\u0026nbsp;23).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003eCombining CSF-based atrophy mapping with other methods\u003c/h2\u003e \u003cp\u003eWe next explored whether CSF-based atrophy mapping could be combined with other atrophy measures to improve performance on any of our four outcome measures. For correlation with neuroradiologist\u0026rsquo;s scores (Supplementary Fig.\u0026nbsp;24), adding CSF to other atrophy methods improved correlations (p\u003csub\u003emax\u003c/sub\u003e = 0.032) but nothing outperformed CSF alone (p\u003csub\u003emin\u003c/sub\u003e = 0.87). For correlation with cognitive scores (Supplementary Fig.\u0026nbsp;25), adding CSF to other atrophy methods improved correlations (p\u0026thinsp;=\u0026thinsp;0.37). Multivariate regression (Supplementary Fig.\u0026nbsp;26) demonstrated the combined CSF\u0026thinsp;+\u0026thinsp;GM atrophy maps explained the most clinical variance in ADAS-Cog scores (R\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.54, p\u0026thinsp;=\u0026thinsp;0.0001), and outperformed CSF alone (DR\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.15, p\u003csub\u003eFWE\u003c/sub\u003e = 0.0001). For localization of verbal memory, adding CSF to other atrophy scores improved the correlation with delayed word recall (Rho\u003csub\u003emin\u003c/sub\u003e = 0.66, p\u003csub\u003emax\u003c/sub\u003e = 0.0000) but did not outperform correlations based on CSF alone (p\u003csub\u003emin\u003c/sub\u003e = 0.11). Finaly, for diagnostic discrimination we found CSF combined with GM provided the overall highest discrimination of control, MCI, and AD patients (AUROC\u0026thinsp;=\u0026thinsp;0.68, p\u0026thinsp;=\u0026thinsp;0.001), but not significantly better than CSF alone (p\u0026thinsp;=\u0026thinsp;0.32).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe found CSF-based atrophy was most consistent with clinical evaluations, which is not surprising as clinical evaluation is based in part on visualizing expanded CSF spaces.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e\u003c/sup\u003e However, we also found that CSF-based atrophy was also most correlated with clinical symptoms, best localized neurological functions, and yielded the highest diagnostic utility. These latter findings are surprising, as CSF expansion is only an indirect proxy of neurodegeneration and lacks the neuroanatomical specificity of techniques like GM or CTh.\u003c/p\u003e \u003cp\u003eThere are good reasons to think that CSF-based metrics would not perform well in detecting clinically relevant brain atrophy, as the majority of neuroimaging research has focused on more direct measures of atrophy based on grey matter, white matter, or cortical thickness.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e However, expansion of CSF spaces remains a hallmark of the clinical approach to identifying atrophy on MRI scans.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e,\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e Our results support this clinical practice and may provide insight into why this clinical approach to atrophy detection has survived the test of time. Neurodegenerative diseases such as AD are known to impact many tissue types simultaneously, including grey mater, white matter, and cortical thickness.\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan additionalcitationids=\"CR46\" citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e,\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e,\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e\u003c/sup\u003e Techniques focused on just one of these tissue types alone may fail to capture clinically relevant atrophy in the other tissue types. While CSF-based atrophy is non-specific, it may do a good job of capturing the summed effect of atrophy across tissue types. Second, the boundary between CSF and non-CSF is easier to detect due to high contrast between CSF and brain. In contrast, it may be harder to detect boundaries between brain tissue types (e.g. GM vs WM). This may render visual or automated methods based on CSF more accurate. Future work is needed to determine why CSF performs well.\u003c/p\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003eRelevance to Aging\u003c/h2\u003e \u003cp\u003eAtrophy can be caused by a number of etiologies, with common etiologies being neurodegeneration and aging. In this manuscript, we compared two different methods of defining brain atrophy, the z-score and w-score. We found the Z-score provides the best overall assessment of atrophy, and this has one critical implication for aging. Z-scores inherently detect age-related atrophy, while W-scores remove age-related atrophy. Thus, CSF-based z-scored atrophy may be a useful tool for aging researchers to better understand how age influences the brain, in derivation of brain-age metrics, or as a useful feature in machine learning algorithms.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003eRelevance to Research\u003c/h2\u003e \u003cp\u003eAutomated atrophy detection methods using GM, WM, and CTh work well and have a longstanding history of successful applications to structural imaging.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan additionalcitationids=\"CR11 CR12 CR13 CR14 CR15\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e However, we found that CSF added additional performance beyond these more common methods, a finding which is consistent with a small number of studies that have examined CSF-based atrophy. Our results confirm and expand this prior work, showing that CSF-based methods outperform other GM, WM, and CTh across 4 different clinically relevant metrics. We also found that CSF-based atrophy might be combined with other techniques to provide additional explanatory power. Our code is openly available (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Calvinwhow/vbm.git\u003c/span\u003e\u003cspan address=\"https://github.com/Calvinwhow/vbm.git\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003eRelevance to Clinical Practice\u003c/h2\u003e \u003cp\u003eCSF-based atrophy correlated well with atrophy assess by a board-certified clinical neuro-radiologist, which may prove useful in ongoing attempts to complement clinical radiology with automated methods.\u003csup\u003e\u003cspan additionalcitationids=\"CR18 CR19 CR20\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e In our study, CSF-based atrophy correlated better with clinical symptoms than atrophy assess by the radiologist. This suggests that our automated method for detecting CSF-based atrophy may detects clinically-relevant atrophy which is not obvious on visual inspection.\u003c/p\u003e \u003cp\u003eOur results suggest that CSF-based atrophy adds additional discriminatory value in distinguishing between AD, MCI, and control patients. This added value was specific to separating MCI from AD, a task which has traditionally been more difficult compared to discriminating patients with AD or MCI from controls.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/div\u003e"},{"header":"Limitations","content":"\u003cp\u003eSeveral limitations apply to this work. First, CSF may not detect atrophy from regions deep within the brain that are not adjacent to CSF spaces. Whether CSF-based atrophy performs well in other disorders outside of MCI/AD, especially those associated with more subcortical atrophy, remains unknown. Second, we performed our neuroimaging analyses in atlas space, which has the potential for registration inaccuracies, especially in the context of brain atrophy.\u003csup\u003e\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e,\u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e\u003c/sup\u003e We used a registration algorithm specifically designed to minimize atrophy-related distortions \u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e but future work is needed to determine the ideal co-registration algorithm or whether automated methods in patient space may perform better. A third limitation is our evaluation of the cerebellum. While the CSF-based atrophy correlated with clinician assessments of cerebellar atrophy, we note the clinicians only had \u0026lsquo;good\u0026rsquo; intraclass correlation in cerebellar evaluation. In part, this is driven by the lack of an existing visual cerebellar atrophy rating scale to act as a gold-standard. Future work will need to develop a gold-standard cerebellar rating scale and relate quantified atrophy to clinician ratings.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eC.W.H. conceptualized the manuscript, collected data, developed all software, performed all analyses, and wrote the manuscript. S.B. assessed clinical scans. E.H. assessed clinical scans. All authors reviewed the manuscript. M.D.F. conceptualized the manuscript, provided supervision, and reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eC.W.H. would like to acknowledge Rachel Bethune Howard.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eAll data is available upon request via ADNI (https://adni.loni.usc.edu/).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eHarper L, Bouwman F, Burton EJ, \u003cem\u003eet al.\u003c/em\u003e Patterns of atrophy in pathologically confirmed dementias: a voxelwise analysis. \u003cem\u003eJ Neurol Neurosurg Psychiatry\u003c/em\u003e 2017; \u003cstrong\u003e88\u003c/strong\u003e: 908\u0026ndash;16.\u003c/li\u003e\n\u003cli\u003eVerhagen MV, Guit GL, Hafkamp GJ, Kalisvaart K. The impact of MRI combined with visual rating scales on the clinical diagnosis of dementia: a prospective study. \u003cem\u003eEur Radiol\u003c/em\u003e 2016; \u003cstrong\u003e26\u003c/strong\u003e: 1716\u0026ndash;22.\u003c/li\u003e\n\u003cli\u003eHarper L, Barkhof F, Fox NC, Schott JM. Using visual rating to diagnose dementia: a critical evaluation of MRI atrophy scales. \u003cem\u003eJ Neurol Neurosurg Psychiatry\u003c/em\u003e 2015; \u003cstrong\u003e86\u003c/strong\u003e: 1225\u0026ndash;33.\u003c/li\u003e\n\u003cli\u003eHarper L, Fumagalli GG, Barkhof F, \u003cem\u003eet al.\u003c/em\u003e MRI visual rating scales in the diagnosis of dementia: evaluation in 184 post-mortem confirmed cases. \u003cem\u003eBrain\u003c/em\u003e 2016; \u003cstrong\u003e139\u003c/strong\u003e: 1211\u0026ndash;25.\u003c/li\u003e\n\u003cli\u003eFox NC, Schott JM. Imaging cerebral atrophy: normal ageing to Alzheimer\u0026rsquo;s disease. \u003cem\u003eThe Lancet\u003c/em\u003e 2004; \u003cstrong\u003e363\u003c/strong\u003e: 392\u0026ndash;4.\u003c/li\u003e\n\u003cli\u003eVirhammar J, Laurell K, Cesarini KG, Larsson E-M. Preoperative Prognostic Value of MRI Findings in 108 Patients with Idiopathic Normal Pressure Hydrocephalus. \u003cem\u003eAm J Neuroradiol\u003c/em\u003e 2014; \u003cstrong\u003e35\u003c/strong\u003e: 2311\u0026ndash;8.\u003c/li\u003e\n\u003cli\u003eRudie JD, Rauschecker AM, Nabavizadeh SA, Mohan S. Neuroimaging of Dilated Perivascular Spaces: From Benign and Pathologic Causes to Mimics. \u003cem\u003eJ Neuroimaging\u003c/em\u003e 2018; \u003cstrong\u003e28\u003c/strong\u003e: 139\u0026ndash;49.\u003c/li\u003e\n\u003cli\u003eDale AM, Fischl B, Sereno MI. Cortical Surface-Based Analysis. \u003cem\u003eNeuroImage\u003c/em\u003e 1999; \u003cstrong\u003e9\u003c/strong\u003e: 179\u0026ndash;94.\u003c/li\u003e\n\u003cli\u003eStatistical Parametric Mapping. Elsevier, 2007 DOI:10.1016/B978-0-12-372560-8.X5000-1.\u003c/li\u003e\n\u003cli\u003eOssenkoppele R, Cohn‐Sheehy BI, La Joie R, \u003cem\u003eet al.\u003c/em\u003e Atrophy patterns in early clinical stages across distinct phenotypes of A lzheimer\u0026rsquo;s disease. \u003cem\u003eHum Brain Mapp\u003c/em\u003e 2015; \u003cstrong\u003e36\u003c/strong\u003e: 4421\u0026ndash;37.\u003c/li\u003e\n\u003cli\u003eTetreault AM, Phan T, Orlando D, \u003cem\u003eet al.\u003c/em\u003e Network localization of clinical, cognitive, and neuropsychiatric symptoms in Alzheimer\u0026rsquo;s disease. \u003cem\u003eBrain\u003c/em\u003e 2020; \u003cstrong\u003e143\u003c/strong\u003e: 1249\u0026ndash;60.\u003c/li\u003e\n\u003cli\u003eHardcastle C, O\u0026rsquo;Shea A, Kraft JN, \u003cem\u003eet al.\u003c/em\u003e Contributions of Hippocampal Volume to Cognition in Healthy Older Adults. \u003cem\u003eFront Aging Neurosci\u003c/em\u003e 2020; \u003cstrong\u003e12\u003c/strong\u003e: 593833.\u003c/li\u003e\n\u003cli\u003eVan Loenhoud AC, Van Der Flier WM, Wink AM, \u003cem\u003eet al.\u003c/em\u003e Cognitive reserve and clinical progression in Alzheimer disease: A paradoxical relationship. \u003cem\u003eNeurology\u003c/em\u003e 2019; \u003cstrong\u003e93\u003c/strong\u003e. DOI:10.1212/WNL.0000000000007821.\u003c/li\u003e\n\u003cli\u003eEgger K, Rau A, Yang S, \u003cem\u003eet al.\u003c/em\u003e Automated voxel- and region-based analysis of gray matter and cerebrospinal fluid space in primary dementia disorders. \u003cem\u003eBrain Res\u003c/em\u003e 2020; \u003cstrong\u003e1739\u003c/strong\u003e: 146800.\u003c/li\u003e\n\u003cli\u003eArbabshirani MR, Plis S, Sui J, Calhoun VD. Single subject prediction of brain disorders in neuroimaging: Promises and pitfalls. \u003cem\u003eNeuroImage\u003c/em\u003e 2017; \u003cstrong\u003e145\u003c/strong\u003e: 137\u0026ndash;65.\u003c/li\u003e\n\u003cli\u003ePark J-Y, Na HK, Kim S, \u003cem\u003eet al.\u003c/em\u003e Robust Identification of Alzheimer\u0026rsquo;s Disease subtypes based on cortical atrophy patterns. \u003cem\u003eSci Rep\u003c/em\u003e 2017; \u003cstrong\u003e7\u003c/strong\u003e: 43270.\u003c/li\u003e\n\u003cli\u003eSeif M, Ziegler G, Freund P. Progressive Ventricles Enlargement and Cerebrospinal Fluid Volume Increases as a Marker of Neurodegeneration in Patients with Spinal Cord Injury: A Longitudinal Magnetic Resonance Imaging Study. \u003cem\u003eJ Neurotrauma\u003c/em\u003e 2018; \u003cstrong\u003e35\u003c/strong\u003e: 2941\u0026ndash;6.\u003c/li\u003e\n\u003cli\u003eChua SE, Cheung C, Cheung V, \u003cem\u003eet al.\u003c/em\u003e Cerebral grey, white matter and csf in never-medicated, first-episode schizophrenia. \u003cem\u003eSchizophr Res\u003c/em\u003e 2007; \u003cstrong\u003e89\u003c/strong\u003e: 12\u0026ndash;21.\u003c/li\u003e\n\u003cli\u003eNyatega CO, Qiang L, Adamu MJ, Kawuwa HB. Gray matter, white matter and cerebrospinal fluid abnormalities in Parkinson\u0026rsquo;s disease: A voxel-based morphometry study. \u003cem\u003eFront Psychiatry\u003c/em\u003e 2022; \u003cstrong\u003e13\u003c/strong\u003e: 1027907.\u003c/li\u003e\n\u003cli\u003eVan Der Kleij LA, De Bresser J, Hendrikse J, Siero JCW, Petersen ET, De Vis JB. Fast CSF MRI for brain segmentation; Cross-validation by comparison with 3D T1-based brain segmentation methods. \u003cem\u003ePLOS ONE\u003c/em\u003e 2018; \u003cstrong\u003e13\u003c/strong\u003e: e0196119.\u003c/li\u003e\n\u003cli\u003eDe Vis JB, Zwanenburg JJ, Van Der Kleij LA, \u003cem\u003eet al.\u003c/em\u003e Cerebrospinal fluid volumetric MRI mapping as a simple measurement for evaluating brain atrophy. \u003cem\u003eEur Radiol\u003c/em\u003e 2016; \u003cstrong\u003e26\u003c/strong\u003e: 1254\u0026ndash;62.\u003c/li\u003e\n\u003cli\u003eDe Vis JB, Zwanenburg JJ, Van Der Kleij LA, \u003cem\u003eet al.\u003c/em\u003e Cerebrospinal fluid volumetric MRI mapping as a simple measurement for evaluating brain atrophy. \u003cem\u003eEur Radiol\u003c/em\u003e 2016; \u003cstrong\u003e26\u003c/strong\u003e: 1254\u0026ndash;62.\u003c/li\u003e\n\u003cli\u003ePetersen RC, Aisen PS, Beckett LA, \u003cem\u003eet al.\u003c/em\u003e Alzheimer\u0026rsquo;s Disease Neuroimaging Initiative (ADNI): Clinical characterization. \u003cem\u003eNeurology\u003c/em\u003e 2010; \u003cstrong\u003e74\u003c/strong\u003e: 201\u0026ndash;9.\u003c/li\u003e\n\u003cli\u003eO\u0026rsquo;Bryant SE, Humphreys JD, Smith GE, \u003cem\u003eet al.\u003c/em\u003e Detecting Dementia With the Mini-Mental State Examination in Highly Educated Individuals. \u003cem\u003eArch Neurol\u003c/em\u003e 2008; \u003cstrong\u003e65\u003c/strong\u003e. DOI:10.1001/archneur.65.7.963.\u003c/li\u003e\n\u003cli\u003eO\u0026rsquo;Bryant SE. Staging Dementia Using Clinical Dementia Rating Scale Sum of Boxes Scores: A Texas Alzheimer\u0026rsquo;s Research Consortium Study. \u003cem\u003eArch Neurol\u003c/em\u003e 2008; \u003cstrong\u003e65\u003c/strong\u003e: 1091.\u003c/li\u003e\n\u003cli\u003eElwood RW. The Wechsler Memory Scale?Revised: Psychometric characteristics and clinical application. \u003cem\u003eNeuropsychol Rev\u003c/em\u003e 1991; \u003cstrong\u003e2\u003c/strong\u003e: 179\u0026ndash;201.\u003c/li\u003e\n\u003cli\u003eAmerican Psychiatric Association. Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition. American Psychiatric Association, 2013 DOI:10.1176/appi.books.9780890425596.\u003c/li\u003e\n\u003cli\u003eWinblad B, Palmer K, Kivipelto M, \u003cem\u003eet al.\u003c/em\u003e Mild cognitive impairment - beyond controversies, towards a consensus: report of the International Working Group on Mild Cognitive Impairment. \u003cem\u003eJ Intern Med\u003c/em\u003e 2004; \u003cstrong\u003e256\u003c/strong\u003e: 240\u0026ndash;6.\u003c/li\u003e\n\u003cli\u003eMcKhann G, Drachman D, Folstein M, Katzman R, Price D, Stadlan EM. Clinical diagnosis of Alzheimer\u0026rsquo;s disease: Report of the NINCDS‐ADRDA Work Group* under the auspices of Department of Health and Human Services Task Force on Alzheimer\u0026rsquo;s Disease. \u003cem\u003eNeurology\u003c/em\u003e 1984; \u003cstrong\u003e34\u003c/strong\u003e: 939\u0026ndash;939.\u003c/li\u003e\n\u003cli\u003eJack CR, Bernstein MA, Fox NC, \u003cem\u003eet al.\u003c/em\u003e The Alzheimer\u0026rsquo;s disease neuroimaging initiative (ADNI): MRI methods. \u003cem\u003eJ Magn Reson Imaging\u003c/em\u003e 2008; \u003cstrong\u003e27\u003c/strong\u003e: 685\u0026ndash;91.\u003c/li\u003e\n\u003cli\u003eScheltens P, Launer LJ, Barkhof F, Weinstein HC, Gool WA. Visual assessment of medial temporal lobe atrophy on magnetic resonance imaging: Interobserver reliability. \u003cem\u003eJ Neurol\u003c/em\u003e 1995; \u003cstrong\u003e242\u003c/strong\u003e: 557\u0026ndash;60.\u003c/li\u003e\n\u003cli\u003eRau A, Urbach H. The MTA score\u0026mdash;simple and reliable, the best for now? \u003cem\u003eEur Radiol\u003c/em\u003e 2021; \u003cstrong\u003e31\u003c/strong\u003e: 9057\u0026ndash;9.\u003c/li\u003e\n\u003cli\u003eClaus JJ, Staekenborg SS, Holl DC, \u003cem\u003eet al.\u003c/em\u003e Practical use of visual medial temporal lobe atrophy cut-off scores in Alzheimer\u0026rsquo;s disease: Validation in a large memory clinic population. \u003cem\u003eEur Radiol\u003c/em\u003e 2017; \u003cstrong\u003e27\u003c/strong\u003e: 3147\u0026ndash;55.\u003c/li\u003e\n\u003cli\u003ePasquier F, Leys D, Weerts JGE, Mounier-Vehier F, Barkhof F, Scheltens P. Inter-and Intraobserver Reproducibility of Cerebral Atrophy Assessment on MRI Scans with Hemispheric Infarcts. \u003cem\u003eEur Neurol\u003c/em\u003e 1996; \u003cstrong\u003e36\u003c/strong\u003e: 268\u0026ndash;72.\u003c/li\u003e\n\u003cli\u003eFrom the Imaging Cognitive Impairment Network (ICINET), Wahlund L-O, Westman E, \u003cem\u003eet al.\u003c/em\u003e Imaging biomarkers of dementia: recommended visual rating scales with teaching cases. \u003cem\u003eInsights Imaging\u003c/em\u003e 2017; \u003cstrong\u003e8\u003c/strong\u003e: 79\u0026ndash;90.\u003c/li\u003e\n\u003cli\u003eGaser C, Dahnke R, Thompson PM, Kurth F, Luders E, Alzheimer\u0026rsquo;s Disease Neuroimaging Initiative. CAT \u0026ndash; A Computational Anatomy Toolbox for the Analysis of Structural MRI Data. Neuroscience, 2022 DOI:10.1101/2022.06.11.495736.\u003c/li\u003e\n\u003cli\u003eManj\u0026oacute;n JV, Coup\u0026eacute; P, Mart\u0026iacute;‐Bonmat\u0026iacute; L, Collins DL, Robles M. Adaptive non‐local means denoising of MR images with spatially varying noise levels. \u003cem\u003eJ Magn Reson Imaging\u003c/em\u003e 2010; \u003cstrong\u003e31\u003c/strong\u003e: 192\u0026ndash;203.\u003c/li\u003e\n\u003cli\u003eAshburner J, Friston KJ. Unified segmentation. \u003cem\u003eNeuroImage\u003c/em\u003e 2005; \u003cstrong\u003e26\u003c/strong\u003e: 839\u0026ndash;51.\u003c/li\u003e\n\u003cli\u003eKlein A, Andersson J, Ardekani BA, \u003cem\u003eet al.\u003c/em\u003e Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. \u003cem\u003eNeuroImage\u003c/em\u003e 2009; \u003cstrong\u003e46\u003c/strong\u003e: 786\u0026ndash;802.\u003c/li\u003e\n\u003cli\u003eAshburner J. A fast diffeomorphic image registration algorithm. \u003cem\u003eNeuroImage\u003c/em\u003e 2007; \u003cstrong\u003e38\u003c/strong\u003e: 95\u0026ndash;113.\u003c/li\u003e\n\u003cli\u003ePereira JMS, Williams GB, Acosta-Cabronero J, \u003cem\u003eet al.\u003c/em\u003e Atrophy patterns in histologic vs clinical groupings of frontotemporal lobar degeneration. \u003cem\u003eNeurology\u003c/em\u003e 2009; \u003cstrong\u003e72\u003c/strong\u003e: 1653\u0026ndash;60.\u003c/li\u003e\n\u003cli\u003eCaspers J, Heeger A, Turowski B, Rubbert C. Automated age- and sex-specific volumetric estimation of regional brain atrophy: workflow and feasibility. \u003cem\u003eEur Radiol\u003c/em\u003e 2021; \u003cstrong\u003e31\u003c/strong\u003e: 1043\u0026ndash;8.\u003c/li\u003e\n\u003cli\u003ePlant C, Teipel SJ, Oswald A, \u003cem\u003eet al.\u003c/em\u003e Automated detection of brain atrophy patterns based on MRI for the prediction of Alzheimer\u0026rsquo;s disease. \u003cem\u003eNeuroImage\u003c/em\u003e 2010; \u003cstrong\u003e50\u003c/strong\u003e: 162\u0026ndash;74.\u003c/li\u003e\n\u003cli\u003eSchultz V, Hedderich DM, Schmitz-Koep B, \u003cem\u003eet al.\u003c/em\u003e Removing outliers from the normative database improves regional atrophy detection in single-subject voxel-based morphometry. \u003cem\u003eNeuroradiology\u003c/em\u003e 2024; \u003cstrong\u003e66\u003c/strong\u003e: 507\u0026ndash;19.\u003c/li\u003e\n\u003cli\u003eShiino A, Watanabe T, Maeda K, Kotani E, Akiguchi I, Matsuda M. Four subgroups of Alzheimer\u0026rsquo;s disease based on patterns of atrophy using VBM and a unique pattern for early onset disease. \u003cem\u003eNeuroImage\u003c/em\u003e 2006; \u003cstrong\u003e33\u003c/strong\u003e: 17\u0026ndash;26.\u003c/li\u003e\n\u003cli\u003eChetelat G, Desgranges B, Landeau B, \u003cem\u003eet al.\u003c/em\u003e Direct voxel-based comparison between grey matter hypometabolism and atrophy in Alzheimer\u0026rsquo;s disease. \u003cem\u003eBrain\u003c/em\u003e 2007; \u003cstrong\u003e131\u003c/strong\u003e: 60\u0026ndash;71.\u003c/li\u003e\n\u003cli\u003eLa Joie R, Perrotin A, Barr\u0026eacute; L, \u003cem\u003eet al.\u003c/em\u003e Region-Specific Hierarchy between Atrophy, Hypometabolism, and \u0026beta;-Amyloid (A\u0026beta;) Load in Alzheimer\u0026rsquo;s Disease Dementia. \u003cem\u003eJ Neurosci\u003c/em\u003e 2012; \u003cstrong\u003e32\u003c/strong\u003e: 16265\u0026ndash;73.\u003c/li\u003e\n\u003cli\u003ePerry DC, Brown JA, Possin KL, \u003cem\u003eet al.\u003c/em\u003e Clinicopathological correlations in behavioural variant frontotemporal dementia. \u003cem\u003eBrain\u003c/em\u003e 2017; \u003cstrong\u003e140\u003c/strong\u003e: 3329\u0026ndash;45.\u003c/li\u003e\n\u003cli\u003eKueper JK, Speechley M, Montero-Odasso M. The Alzheimer\u0026rsquo;s Disease Assessment Scale\u0026ndash;Cognitive Subscale (ADAS-Cog): Modifications and Responsiveness in Pre-Dementia Populations. A Narrative Review. \u003cem\u003eJ Alzheimers Dis\u003c/em\u003e 2018; \u003cstrong\u003e63\u003c/strong\u003e: 423\u0026ndash;44.\u003c/li\u003e\n\u003cli\u003eMa J, Zheng M-X, Wu J-J, \u003cem\u003eet al.\u003c/em\u003e Mapping the long-term delayed recall-based cortex-hippocampus network constrained by the structural and functional connectome: a case-control multimodal MRI study. \u003cem\u003eAlzheimers Res Ther\u003c/em\u003e 2023; \u003cstrong\u003e15\u003c/strong\u003e: 61.\u003c/li\u003e\n\u003cli\u003eFoster JK, Meikle A, Goodson G, \u003cem\u003eet al.\u003c/em\u003e The Hippocampus and Delayed Recall: Bigger is not Necessarily Better? \u003cem\u003eMemory\u003c/em\u003e 1999; \u003cstrong\u003e7\u003c/strong\u003e: 715\u0026ndash;33.\u003c/li\u003e\n\u003cli\u003eKramer JH, Schuff N, Reed BR, \u003cem\u003eet al.\u003c/em\u003e Hippocampal volume and retention in Alzheimer\u0026rsquo;s disease. \u003cem\u003eJ Int Neuropsychol Soc\u003c/em\u003e 2004; \u003cstrong\u003e10\u003c/strong\u003e: 639\u0026ndash;43.\u003c/li\u003e\n\u003cli\u003eMa J, Wu J-J, Xing X-X, \u003cem\u003eet al.\u003c/em\u003e Circuit-based neuromodulation enhances delayed recall in amnestic mild cognitive impairment. \u003cem\u003eJ Neurol Neurosurg Psychiatry\u003c/em\u003e 2024; : jnnp-2023-333152.\u003c/li\u003e\n\u003cli\u003eMortimer JA, Gosche KM, Riley KP, Markesbery WR, Snowdon DA. Delayed recall, hippocampal volume and Alzheimer neuropathology: Findings from the Nun Study. \u003cem\u003eNeurology\u003c/em\u003e 2004; \u003cstrong\u003e62\u003c/strong\u003e: 428\u0026ndash;32.\u003c/li\u003e\n\u003cli\u003eWinkler AM, Ridgway GR, Webster MA, Smith SM, Nichols TE. Permutation inference for the general linear model. \u003cem\u003eNeuroImage\u003c/em\u003e 2014; \u003cstrong\u003e92\u003c/strong\u003e: 381\u0026ndash;97.\u003c/li\u003e\n\u003cli\u003eRaschka S. Model Evaluation, Model Selection, and Algorithm Selection in Machine Learning. 2018. DOI:10.48550/ARXIV.1811.12808.\u003c/li\u003e\n\u003cli\u003eWhang Y-J. Consistent bootstrap tests of parametric regression functions. \u003cem\u003eJ Econom\u003c/em\u003e 2000; \u003cstrong\u003e98\u003c/strong\u003e: 27\u0026ndash;46.\u003c/li\u003e\n\u003cli\u003eBarnett AJ, Guo Z, Jing J, \u003cem\u003eet al.\u003c/em\u003e Improving Clinician Performance in Classifying EEG Patterns on the Ictal\u0026ndash;Interictal Injury Continuum Using Interpretable Machine Learning. \u003cem\u003eNEJM AI\u003c/em\u003e 2024; \u003cstrong\u003e1\u003c/strong\u003e. DOI:10.1056/AIoa2300331.\u003c/li\u003e\n\u003cli\u003eSeabold S, Perktold J. Statsmodels: Econometric and Statistical Modeling with Python. In: Proceedings of the 9th Python in Science Conference. 2010: 92\u0026ndash;6.\u003c/li\u003e\n\u003cli\u003eAbraham A, Pedregosa F, Eickenberg M, \u003cem\u003eet al.\u003c/em\u003e Machine learning for neuroimaging with scikit-learn. \u003cem\u003eFront Neuroinformatics\u003c/em\u003e 2014; \u003cstrong\u003e8\u003c/strong\u003e. DOI:10.3389/fninf.2014.00014.\u003c/li\u003e\n\u003cli\u003ePedregosa F, Varoquaux G, Gramfort A, \u003cem\u003eet al.\u003c/em\u003e Scikit-learn: Machine Learning in Python. \u003cem\u003eJ Mach Learn Res\u003c/em\u003e 2011; \u003cstrong\u003e12\u003c/strong\u003e: 2825\u0026ndash;30.\u003c/li\u003e\n\u003cli\u003eVargha-Khadem F, Gadian DG, Watkins KE, Connelly A, Van Paesschen W, Mishkin M. Differential Effects of Early Hippocampal Pathology on Episodic and Semantic Memory. \u003cem\u003eScience\u003c/em\u003e 1997; \u003cstrong\u003e277\u003c/strong\u003e: 376\u0026ndash;80.\u003c/li\u003e\n\u003cli\u003eYoshii F, Kawaguchi C, Kohara S, \u003cem\u003eet al.\u003c/em\u003e Characteristic deterioration of ADAS-Jcog subscale scores and correlations with regional cerebral blood flow reductions in Alzheimer\u0026rsquo;s disease. \u003cem\u003eNeurol Sci\u003c/em\u003e 2018; \u003cstrong\u003e39\u003c/strong\u003e: 909\u0026ndash;18.\u003c/li\u003e\n\u003cli\u003eon behalf of the MAGNIMS study group, Sastre-Garriga J, Pareto D, \u003cem\u003eet al.\u003c/em\u003e MAGNIMS consensus recommendations on the use of brain and spinal cord atrophy measures in clinical practice. \u003cem\u003eNat Rev Neurol\u003c/em\u003e 2020; \u003cstrong\u003e16\u003c/strong\u003e: 171\u0026ndash;82.\u003c/li\u003e\n\u003cli\u003eHobden G, Colbourne E, Pendlebury ST, Demeyere N. Reliability of the global cortical atrophy visual rating scale applied to computed tomography versus magnetic resonance imaging scans in acute stroke. \u003cem\u003eNeurol Sci\u003c/em\u003e 2024; \u003cstrong\u003e45\u003c/strong\u003e: 1549\u0026ndash;56.\u003c/li\u003e\n\u003cli\u003eCardinale F, Chinnici G, Bramerio M, \u003cem\u003eet al.\u003c/em\u003e Validation of FreeSurfer-Estimated Brain Cortical Thickness: Comparison with Histologic Measurements. \u003cem\u003eNeuroinformatics\u003c/em\u003e 2014; \u003cstrong\u003e12\u003c/strong\u003e: 535\u0026ndash;42.\u003c/li\u003e\n\u003cli\u003eYao Z, Hu B, Liang C, Zhao L, Jackson M, the Alzheimer\u0026rsquo;s Disease Neuroimaging Initiative. A Longitudinal Study of Atrophy in Amnestic Mild Cognitive Impairment and Normal Aging Revealed by Cortical Thickness. \u003cem\u003ePLoS ONE\u003c/em\u003e 2012; \u003cstrong\u003e7\u003c/strong\u003e: e48973.\u003c/li\u003e\n\u003cli\u003eOxenford S, R\u0026iacute;os AS, Hollunder B, \u003cem\u003eet al.\u003c/em\u003e WarpDrive: Improving spatial normalization using manual refinements. \u003cem\u003eMed Image Anal\u003c/em\u003e 2024; \u003cstrong\u003e91\u003c/strong\u003e: 103041.\u003c/li\u003e\n\u003cli\u003ePereira JMS, Xiong L, Acosta-Cabronero J, Pengas G, Williams GB, Nestor PJ. Registration accuracy for VBM studies varies according to region and degenerative disease grouping. \u003cem\u003eNeuroImage\u003c/em\u003e 2010; \u003cstrong\u003e49\u003c/strong\u003e: 2205\u0026ndash;15.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6423862/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6423862/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eClinical neuroradiologists routinely look for expansion of CSF spaces to help identify atrophy on patient MRI scans. In contrast, automated methods for identifying atrophy rely on changes in grey matter volume or cortical thickness. It is unclear if evaluating CSF spaces could improve detection of brain atrophy, which may be relevant to improving detection of age- and disease-related atrophy.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003e3 clinician experts graded atrophy across 7 brain regions from 50 subjects enrolled in the Alzheimer Disease Neuroimaging Initiative (Discovery Cohort, n\u0026thinsp;=\u0026thinsp;1050 visual ratings) while one expert graded atrophy in an additional 150 subjects (Validation cohort, n\u0026thinsp;=\u0026thinsp;1050 visual ratings). These subjects included patients with mild cognitive impairment (MCI, n\u0026thinsp;=\u0026thinsp;72), Alzheimer\u0026rsquo;s disease (AD, n\u0026thinsp;=\u0026thinsp;60), and age-matched healthy controls (n\u0026thinsp;=\u0026thinsp;68), randomly selected from the broader sample. We used an automated approach to detect expansion of CSF spaces and compared it with standard methods for detecting brain atrophy (grey/white matter volume, cortical thickness). We evaluated four metrics of performance: 1) correlation to visually rated atrophy; 2) correlation to clinical symptoms; 3) localization of atrophy most correlated with verbal memory scores; and 4) ability to discriminate between AD, MCI, and controls.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eAtrophy detected by expansion of CSF spaces significantly outperformed existing methods across all performance metrics. 1) CSF-based atrophy correlated with manually assessed atrophy scores (Median Rho\u0026thinsp;=\u0026thinsp;0.50, p\u003csub\u003emax\u003c/sub\u003e = 0.043), and this correlation was stronger than all other methods (max p\u003csub\u003eFWE\u003c/sub\u003e = 0.0005). 2) CSF-based atrophy correlated with clinical symptoms (Median Rho\u0026thinsp;=\u0026thinsp;0.37, IQR 0.34\u0026ndash;0.46), and this correlation was stronger than all other methods (max p\u003csub\u003eFWE\u003c/sub\u003e = 0.0015). 3) CSF-based atrophy was the only method to localize FWE-significant atrophy covarying with verbal memory scores to the left hippocampus (Rho\u0026thinsp;=\u0026thinsp;0.57, p\u003csub\u003eFWE\u003c/sub\u003e = 0.00302). 4) CSF-based atrophy best differentiated between AD, MCI, and controls (AUC\u0026thinsp;=\u0026thinsp;0.68, 95% CI 0.61\u0026ndash;0.75), and outperformed all other methods (max p\u003csub\u003eFWE\u003c/sub\u003e = 0.041). All results were reproducible across discovery and replication cohorts.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eDeriving brain atrophy using CSF can increase sensitivity of atrophy detection, improving alignment with clinical evaluations, explained variance, localization strength, and diagnostic utility.\u003c/p\u003e","manuscriptTitle":"Using Cerebrospinal Fluid Improves Detection of Individual Brain Atrophy","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-07 16:50:36","doi":"10.21203/rs.3.rs-6423862/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"81f15265-2d3e-4cb9-80f1-bd3c4caeb535","owner":[],"postedDate":"May 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":47912794,"name":"Biological sciences/Neuroscience"},{"id":47912795,"name":"Health sciences/Anatomy/Nervous system"}],"tags":[],"updatedAt":"2025-05-29T14:24:02+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-07 16:50:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6423862","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6423862","identity":"rs-6423862","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}