Artificial intelligence for volumetric measurement of cerebral white matter hyperintensities on thick-slice fluid-attenuated inversion recovery (FLAIR) magnetic resonance images from multiple centers

Artificial intelligence for volumetric measurement of cerebral white matter hyperintensities on thick-slice fluid-attenuated inversion recovery (FLAIR) magnetic resonance images from multiple centers | 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 Artificial intelligence for volumetric measurement of cerebral white matter hyperintensities on thick-slice fluid-attenuated inversion recovery (FLAIR) magnetic resonance images from multiple centers Masashi Kuwabara, Fusao Ikawa, Shinji Nakazawa, Saori Koshino, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3833822/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 02 May, 2024 Read the published version in Scientific Reports → Version 1 posted 10 You are reading this latest preprint version Abstract We aimed to develop a new artificial intelligence software that can automatically extract and measure the volume of white matter hyperintensities (WMHs) in head magnetic resonance imaging (MRI) using only thick-slice fluid-attenuated inversion recovery (FLAIR) sequences from multiple centers. We enrolled 1,092 participants in Japan, comprising this thick-slice Private Dataset. Based on 207 randomly selected participants, neuroradiologists annotated WMHs using predefined guidelines. The annotated participants were divided into training (n = 138) and test (n = 69) datasets. The WMH segmentation model comprised a U-Net ensemble and was trained using the Private Dataset. Two other models were trained for validation using either both thin- and thick-slice MRI datasets or the thin-slice dataset alone. The voxel-wise Dice similarity coefficient (DSC) was used as the evaluation metric. The model trained using only thick-slice MRI showed a DSC of 0.820 for the test dataset, which is comparable to the accuracy of human readers. The model trained with the additional thin-slice dataset showed only a slightly improved DSC of 0.822. This automatic WMH segmentation model comprising a U-Net ensemble trained on a thick-slice FLAIR MRI dataset is a promising new method. Despite some limitations, this model may be applicable in clinical practice. Biological sciences/Neuroscience/Cognitive ageing Physical sciences/Mathematics and computing/Software Health sciences/Anatomy/Nervous system Health sciences/Health care/Medical imaging/Magnetic resonance imaging Biological sciences/Neuroscience Health sciences/Medical research Health sciences/Neurology Physical sciences/Mathematics and computing Figures Figure 1 Figure 2 Figure 3 Introduction White matter hyperintensities (WMHs) are common neuroimaging findings characterized by high signals either as periventricular hyperintensities (PVHs) or deep subcortical WMHs (DSWMHs) in fluid-attenuated inversion recovery (FLAIR) sequences of head magnetic resonance imaging (MRI) [ 1 ]. Many reports have shown associations of the WMH degree with ischemic stroke, depression, dementia, and emotional disorders [ 2 – 8 ]. Although the WMH prevalence is higher in older individuals, studies of healthy younger people have reported that the WMH degree is also correlated with the execution of speed-demanding functions, such as word recall, that is subfrontal cortical functions [ 6 – 10 ]. Brain Dock is a widely implemented brain checkup system that uses MRI and magnetic resonance angiography to detect asymptomatic brain diseases, such as unruptured cerebral aneurysms and WMHs at an early stage, thereby preventing stroke and dementia in healthy individuals in Japan [ 11 , 12 ]. During a Brain Dock, the evaluation of WMHs is required, using a standard database with the aim of accumulating data on the diagnosis and early detection of disease risk [ 11 – 13 ]. To date, manual interpretation by neuroradiologists has been considered an effective method for quantitatively assessing WMHs [ 14 ]. However, this approach is time-consuming, labor-intensive, and has the disadvantage of being unsuitable for clinical application or large-scale studies because of the large interrater variability ranging from 10–68% [ 14 – 17 ]. Recently, artificial intelligence (AI) has been used for automated reading and volumetric measurements. Although many publications have reported the use of AI to assess brain atrophy on T1-weighted images (T1WIs), reports on AI algorithms for automated WMH segmentation using only FLAIR images are limited [ 15 , 18 ]. The size, number, shape, and location of WMHs are heterogeneous and age-dependent, making them difficult to evaluate using AI. Moreover, WMHs are present in a variety of diseases, requiring background information, such as patient history and disease course, and multiple MRI sequences for differentiation [ 15 , 19 , 20 ]. Only a few of the reported AI algorithms are applicable in clinical practice because many are limited to specific diseases, have small sample sizes, evaluate multiple sequences, or require a thin slice thickness as the imaging conditions [ 19 , 21 , 22 ]. To ensure applicability in clinical practice, accurate results must be provided consistently under imaging protocols for various WMHs, often requiring multicenter collaborative studies [ 23 , 24 ]. Only a few studies to date have validated the AI algorithm for automatic WMH segmentation by FLAIR alone in a multicenter setting because of the difficulty [ 25 ]. If AI algorithms for automated WMH segmentation using only FLAIR images were developed and automatically classified in Brain Dock, the scientific contribution to preventive medicine in healthy people would be immeasurable. The initial purpose of this study was to develop a new AI software that can automatically extract and measure WMH volumes on head MRI using only thick-slice FLAIR images from multiple centers. In the future, we plan to develop an AI that can automatically classify WMHs. Results According to the five neuroradiologists participating in this study, the average WMH volume was 18.1 ± 22.1 mL, with a minimum of 0.0 mL and a maximum of 109.5 mL. The distribution of WMH volumes in the Private Dataset (PR) (Fig. 1 a) closely resembled that in the WMH Segmentation Challenge (WMHC) Dataset (Fig. 1 b). Given the potential for biased evaluation data to negatively impact generalization performance on unseen data, the PR was expected to evaluate generalization performance with similar accuracy to that of the WMHC dataset. Moreover, the 885 unannotated participants were utilized for pseudo-labeling to enhance model performance. Quantitative analysis Table 1 shows the evaluation results of WMH segmentation based on the Dice similarity coefficient (DSC), Recall, precision, and modified Hausdorff distance (95th percentile of the Hausdorff distance in mm) were compared. The PR model demonstrated high performance with a DSC of 0.8 or higher on the PR, whereas its performance on the WMHC Dataset was not as high. The PR + WMHC model, which included the WMHC Dataset for training, showed improved performance and the ability to handle both thick- and thin-slice MRI datasets simultaneously. However, its performance on the PR improved only marginally, with a DSC increase of only 0.002. This modest degree of improvement might be attributed to domain differences between these datasets, suggesting that the thin-slice dataset does not significantly enhance performance on the thick-slice dataset. Additionally, the model trained only on the WMHC Dataset exhibited overfitting to the test dataset. Table 1 Assessment of WMH segmentation by the U-Net ensemble models Model name Training dataset Private Dataset for the test dataset (n = 69) WMHC Dataset for the test dataset (n = 110) DSC Recall Precision H95 DSC Recall Precision H95 PR Private Dataset 0.820 0.788 0.855 14.50 0.742 0.862 0.653 22.78 PR + WMHC Private and WMHC Datasets 0.822 0.789 0.858 14.92 0.826 0.877 0.781 7.97 WMHC WMHC Dataset 0.676 0.531 0.931 16.20 0.833 0.874 0.796 5.61 DSC: Dice similarity coefficient, H95: modified Hausdorff distance (95th percentile [mm]), PR: Private Dataset, WMH: white matter hyperintensity, WMHC: white matter hyperintensity segmentation challenge. Qualitative analysis Figure 2 shows the segmentation results of the PR and WMHC Dataset models used with the test dataset. The top row presents a participant with a relatively small WMH volume. All models successfully detected regions that were similar to the ground truth. In a participant with extensive WMHs, as in the second row, both the PR and PR + WMHC models accurately detected the WMH regions, whereas the WMHC model exhibited false negatives (FNs), leading to a lower DSC. However, in a participant with small punctate WMHs, as observed in the right hemisphere in the third row, the WMHC model demonstrated better results compared with the other two models. The differences between the PR and the PR + WMHC models are shown in the last row. False positives (FPs), which occurred in the PR model, were reduced in the PR + WMHC model. For all other participants, the difference between these two models was minimal, which resulted in a slight DSC difference of 0.002. A tendency for FPs and FNs was observed in the PR model in areas with small punctate WMHs, near the lateral ventricles, near the septum pellucidum, lacunar infarcts, or where the boundary between WMHs and other structures was ambiguous. As shown in Fig. 2 , row 3, the model missed the small punctate WMHs in the left hemisphere. Conversely, in the example from Fig. 2 , row 4, punctate FPs can be observed in the left hemisphere. The figure also shows FPs near the surface of the lateral ventricle body, a region where the boundary between a WMH and other structures could be indistinct. While the examples in Fig. 2 , rows 1 and 2, are generally accurate, an unintended “hole” appeared in the right hemisphere of Fig. 2 , row 2. Furthermore, the recall of the PR model was lower than its precision, indicating that the model tended to have an under-detection bias on the PR. To better understand variations on a case-by-case basis, the DSC for each case was calculated using the PR model results of the PR. The DSC values ranged from a minimum of 0.158 to a maximum of 0.901. Figure 3 a and 3 b show the cases with the lowest and highest DSCs, respectively. The case with the lowest DSC, at 0.158, missed small punctate WMHs, and the missed area accounted for a large portion of the total WMH area. Conversely, the case with the highest DSC (0.901) showed a larger total WMH area. Although some FNs or FPs were observed, their proportions to the total count were limited, resulting in a high DSC. Processing time analysis Processing time measurements were conducted for thick-slice MRI images, revealing an average processing time of 18.5 s per case. When corrected for comparison with a previous study, the equivalent value was approximately 1.8 min. Discussion In this study, we evaluated the performance of an AI using automatically extracted volumetric measurement of WMH lesions on head MRI images on thick-slice FLAIR images. The results showed that the DSC in our study reached 0.820, which is comparable to human reading accuracy. In recent years, although various algorithms for AI-based automated volume measurements of WMHs have been reported in the literature, very few are applicable in clinical practice [ 19 , 21 , 22 ]. In previous studies on AI and volumetric measurements of WMHs, many were limited to specific diseases, such as Alzheimer’s disease or multiple sclerosis, and had relatively small sample sizes of approximately 100 patients [ 26 – 28 ] and an algorithms that used three-dimensional (3D) networks with thin-slice image data with 1–3 mm slice thicknesses [ 27 , 29 , 30 ]. Thin-slice MRI has only been widely applied in a few developed countries and is currently not commonly used in clinical practice [ 14 ]. A qualitative evaluation with subjective judgments should be objectively compared with a quantitative evaluation based on automatic volumetric measurement using AI. WMHs result primarily from aging processes, such as demyelination and axonal loss, both of which occur as a result of small vessel disease of the brain [ 7 ]. Thus, WMHs are frequently observed in older asymptomatic patients, although they can also occur in a variety of other diseases, including demyelinating diseases such as multiple sclerosis; genetic diseases such as cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy; infectious diseases such as human immunodeficiency virus encephalopathy; psychiatric disorders such as depressive disorders and schizophrenia; and various other diseases such as autoimmune diseases, neoplastic diseases, intoxication, and hypoxic encephalopathies [ 31 ]. Therefore, volumetric detection of WMHs in younger people is an important aspect of brain checkups, and lifestyle improvements are needed in examinees with increased WMH volume over time. This study proposes a new AI software that solves these problems, and its major feature is 5-mm thick-slice FLAIR images. Moreover, this study evaluated both the PR and the WMHC Dataset as thick (5 mm) and thin (1–3 mm) slices, respectively, with a sample size of over 200 participants. The results showed that the WMH segmentation accuracy based on two-dimensional [2D] FLAIR images with 5 mm slices, which is widely used in general hospitals, was comparable to that of previous studies. This suggests that the proposed AI algorithm can be applied in clinical practice. We only identified 39 publications that analyzed the automatic segmentation of WMHs on head MRI using AI similar to the present study. Of these, we reviewed 17 publications since 2018, including the present study (Table 2 ). The sample size varied from 20 to 1,092 overall, with the sample size of this study being the largest. Multiple MRI sequences analyzing both T1WI and FLAIR images were identified in 12 studies [ 20 , 22 , 26 , 32 – 37 ], and only 1 study in addition to the present study evaluated only FLAIR images. Regarding the MRI field strength, nearly all studies, including ours, used 3T MRI. Regarding the segmentation accuracy, the DSC values ranged from 0.41 to 0.83, showing a wide range of accuracy. The DSC of the AI model in this study was 0.820, indicating higher accuracy compared with that of previous studies overall. Therefore, the model developed in the present study demonstrably has a better performance than that of existing AI models. Table 2 Review of recent studies on the automatic segmentation of cerebral white matter hyperintensities in head MRI using AI Author / Year Sample size Breakdown of samples MRI Type of segmentation Average spatial agreement with reference segmentation Sequence Field strength (T) Manjón JV / 2018 128 N/A T1WI, FLAIR 1.5, 3 Deep learning DSC: 0.78 Knight J / 2018 96 N/A FLAIR 1.5, 3 Supervised DSC: 0.41–0.70 Qin C / 2018 88 N/A T1WI, FLAIR 3 Deep learning DSC: 0.67 Ling Y / 2018 156 N/A T1WI, FLAIR 1.5, 3 Supervised DSC: 0.76 Park BY / 2018 148 N/A T1WI, FLAIR 1.5 Supervised DSC: 0.65 − 0.67 Atlason HE / 2019 170 train: 60, test: 110 T1WI, T2WI, FLAIR 1.5, 3 Unsupervised DSC: 0.53 − 0.67, TPR: 0.25–0.40 Sundaresan V / 2019 133 N/A T1WI, T2WI, FLAIR 1.5, 3 Supervised DSC: 0.77, TPR: 0.73–0.98 Wu D / 2019 135 train: 15, test: 120 T1WI, FLAIR 3 Supervised DSC: 0.62, FPR: 0.35, FNR: 0.37 Wu J / 2019 60 N/A T1WI, FLAIR 1.5, 3 Deep learning DSC: 0.78 Ding T / 2020 20 train: 15, test: 5 T1WI, FLAIR 3 Supervised DSC: 0.78, TPR: 0.70 Fiford CM / 2020 80 train: 20, test: 60 T1WI, FLAIR 3 Unsupervised DSC: 0.74 Hong J / 2020 148 N/A T1WI, FLAIR 3 Deep learning TPR: 0.87, FDR: 0.10 Liu L / 2020 60 test: 60 T1WI, FLAIR 1.5, 3 Deep learning DSC: 0.83 Rachmadi MF / 2020 60 test: 60 T1WI, T2WI, FLAIR 3 Unsupervised DSC: 0.47 − 0.56, TPR: 0.47 Tran P / 2022 60 N/A T1WI, FLAIR N/A N/A DSC: 0.67 Zhu W / 2022 1,045 cross-validation: 849 T1WI, FLAIR 1.5, 3 Supervised DSC: 0.792 Present study / 2023 1,092 train: 138 + 885, test: 69 FLAIR 1.5, 3 Supervised DSC: 0.820 DSC: dice score coefficient, TPR: true positive rate, FPR: false positive rate, FNR: false negative rate, FDR: false discovery rate, FLAIR: fluid-attenuated inversion recovery, T1WI: T1-weighted imaging, T2WI: T2-weighted imaging, T: tesla, N/A: not applicable, MRI: magnetic resonance imaging, AI: artificial intelligence The unique feature of our study is that we have developed an AI that performs better than previous AI models using “thick slices” and “FLAIR images only” in accordance with the standards of the Japan Brain Dock Society. Notably, the performance of this AI model was tested on the largest sample size in a study of this type of more than 1,000 cases, using data from multiple institutions. In addition, the processing time when performing segmentation was as fast as 18.5 s per case for our AI model. When corrected for comparison with a previous study, the equivalent value was approximately 1.8 min. This is faster than the approximate processing time of 5 min using a supervised deep learning-based method reported in a previous study [ 32 ]. Therefore, this AI algorithm seems more appropriate for real-world clinical adaptation. MRI imaging with thick slices and FLAIR images only are simple methods for brain checkups, and thus has an advantage in that MRI examinations are easily accessible and versatile for patients. Our findings can also be used as a basis for PVH and DSWMH grading as proposed by the Japan Brain Dock Society. The ultimate goal of our AI model is to classify PVH and DSWMH grading to screen normal participants at risk of early-onset dementia to prevent or delay the onset of dementia through lifestyle guidance. This study had several limitations. First, approximately 10 physicians diagnosed the images for annotation, possibly leading to individual differences in physician judgment, resulting in diagnostic bias. However, all physicians in this study were board certified radiologists from the Japan Radiological Society or neuroradiologists with more than 7 years of experience and board-certified neurosurgeons from the Japan Neurosurgical Society; therefore, the diagnostic results should be of a uniformly high standard. Second, because only high intensity areas on FLAIR images were evaluated in this study, lacunar infarction might not have been clearly distinguished because only two lacunar infarctions were included in the training dataset. We would like to develop another AI to distinguish between WMHs and lacunar infarctions after acquiring a lacunar infarctions dataset in the near future. Third, a detailed analysis differentiating between PVH and DSWMH within WMHs was not conducted in this study; however, our research is currently ongoing. In conclusion, the automatic WMH segmentation model based on a U-Net ensemble trained on a thick-slice FLAIR MRI dataset showed a DSC of 0.820 and a promising AI algorithm. This model may be applicable in clinical practice for brain checkups. Methods Ethics approval and informed consent This study was approved by the Institutional Review Board of Hiroshima University (approval number: E2022-0262). All study protocols were developed according to the guidelines of Hiroshima University Hospital. Individual data were anonymized and collected during routine Brain Dock examinations; thus, all participants provided informed consent based on an opt-out method. Moreover, in this study, we were provided with anonymized data; therefore, the authors did not have access to personally identifiable participant information during or after data collection. Study design In this study, a new automatic WMH segmentation model was developed based on existing commercial software (EIRL Brain Metry version 1.13.0, LPIXEL Inc., Tokyo, Japan). This software employs U-Net for segmentation. Although the core usage of U-Net remains consistent, we replaced the backbone network with a new model and also explored the model ensembles. The WMH segmentation model of the software was updated using a newly collected dataset, the PR. This new model was named “PR” as it was trained using the PR. The model performance was also evaluated using the publicly available WMHC Dataset [ 29 ]. The WMHC Dataset has thinner slices than those in the PR. For validation, two additional models were trained using the same training procedure. One was a model trained using both the PR and the WMHC Dataset (PR + WMHC model) to evaluate the effect of adding an MRI dataset with thin slices to the PR with thick slices. The other was a model trained using only the WMHC Dataset (the WMHC model) to evaluate the performance of a model trained using only a thin-slice MRI dataset. Private Dataset For this dataset, 1,092 FLAIR MRI images were collected from three clinics in Japan. To minimize potential biases related to the WMH severity, data was gathered in a manner that ensured a balanced distribution of PVH/DSWMH grades (0–4) as determined by each clinic. The WMH annotation process was restricted to 207 randomly selected participants because of the annotation costs. Nearly all the annotated WMHs were age-related, and 2 of these 207 participants had lacunar infarcts. The slice thickness of the annotated data ranged from 5–6 mm, with an average of 5.37 mm for the entire dataset. Detailed spatial characteristics, including slice thickness, are presented in Table 3 . The data was acquired using nine different types of scanners, and the detailed MRI parameters are listed in Supplementary Table S1 . Table 3 Characteristics of the Private Dataset comprising 207 annotated participants Vendor Scanner type ID a Scanner model Magnetic field strength (T) Voxel size (mm 3 ) Training Test GE Healthcare 1 DISCOVERY MR750w 3 6.0 × 0.43 × 0.43 32 0 (Chicago, IL, USA) 2 Optima MR450w 1.5 5.0 × 0.469 × 0.469 0 2 3 SIGNA Architect 3 5.0 × 0.43 × 0.43 3 4 4 SIGNA EXCITE 1.5 6.0 × 0.469 × 0.469 0 6 5 Signa HDxt 1.5 5.0 × 0.469 × 0.469 0 1 5 Signa HDxt 1.5 6.0 × 0.43 × 0.43 1 0 5 Signa HDxt 1.5 6.0 × 0.469 × 0.469 0 1 5 Signa HDxt 1.5 6.0 × 0.859 × 0.859 37 0 6 Signa HDxt 3 5.0 × 0.43 × 0.43 6 11 Philips 7 Ingenia 1.5 5.0 × 0.479 × 0.479 19 8 (Amsterdam, the Netherlands) 7 Ingenia 1.5 5.0 × 0.448 × 0.448 1 0 7 Ingenia 1.5 5.0 × 0.449 × 0.449 8 25 Siemens Healthineers 8 Avanto 1.5 5.0 × 0.898 × 0.898 15 5 (Erlangen, Germany) 9 Symphony 1.5 5.0 × 0.719 × 0.719 16 6 a Scanner type ID: a unique ID for each combination of scanner and magnetic field strength Five neuroradiologists shared the annotation efforts of WMHs for the 207 selected participants. To minimize interrater variability and ensure annotation quality, an annotation guideline was developed under the supervision of neuroradiologists. The following guidelines were provided for areas prone to individual variations: (a) the area around the ventricles, sulci, and longitudinal fissure may show a non-WMH signal; therefore, an annotation is only made when a WMH is clearly visible on anterior to posterior slices, (b) the septum pellucidum of the lateral ventricles should not be annotated, (c) only WMHs in the cerebrum should be annotated, and (d) lacunar infarcts and EPVSs should be excluded. The imaging differentiation between lacunar infarcts, EPVSs, and WMHs is that a WMH is depicted as a clear hyperintense lesion on FLAIR, whereas an EPVS is depicted as an iso-to-hypointense lesion and a lacunar infarct is depicted as an iso-to-hyperintense lesion with a hypointense center [ 11 , 13 ]. In addition, WMHs are found mainly in the cerebral white matter, whereas EPVSs and lacunar infarcts are found in the basal ganglia, in addition to the cerebral white matter. For the model development, the annotated datasets were randomly divided into training (n = 138) and test (n = 69) datasets. To ensure an accurate evaluation, a second annotation review was conducted on 69 test participants. The review was performed by three neuroradiologists who differed from the five physicians employed in the initial annotation process. The reviewers were asked to modify the previously annotated WMH regions if needed. Consequently, some modifications were made to the test dataset. WMH Segmentation Challenge Dataset For the evaluation using an external dataset, the WMHC Dataset was employed [ 29 ]. A competition using this dataset was originally held at the 20th International Conference on Medical Image Computing and Computer-Assisted Intervention in 2017 in Quebec City, Quebec, Canada. After the competition, models could be submitted via their website until December 2022. Afterward, the previously undisclosed test dataset was made available. Several previous studies have used this dataset [ 15 , 26 , 31 , 32 ], and our study assessed performance using the same dataset. Although the dataset included both FLAIR and T1W images, only FLAIR images were used in this study. The WMHC dataset consisted of 60 training and 110 test participants. These participants were sourced from three institutes located in the Netherlands and Singapore. The data was acquired using five different types of scanners. The average slice thickness of the FLAIR sequences in this dataset was 2.23 mm. Notably the average slice thickness was approximately 41% that of the PR. The WMH volume averaged 16.9 ± 21.6 mL, with a minimum of 0.78 mL and a maximum of 195.15 mL [ 29 ]. WMH segmentation model The algorithm for WMH detection in the software comprised a 2D convolutional neural network (CNN) model that processes individual MRI slices sequentially. Initially, pixel values of the input MRI images were normalized to fall within the range of 0.0–1.0. Notably, image processing-based bias field corrections were not applied. We made this decision because the effect of the contrast augmentation was anticipated to accommodate such variations. Then, the normalized image was resized to 512 × 512 pixels and input into the main CNN model. This CNN model was an ensemble of two U-Net models, each with distinct characteristics. The first U-Net model utilized ResNext50 as the encoder, and the ImageNet pretrained weights were used as the initial weights. The learning rate was set to 0.001, the batch size was set to 15, and the model was trained for 15 epochs using the Adam optimizer. These hyperparameters were obtained using a grid search. During training, the learning rate was decayed using a cosine-annealing learning-rate reducer to facilitate convergence. Matthews correlation coefficient loss was used as the loss function, as originally proposed for skin lesion segmentation [ 38 ]. This loss function incorporates a penalty for misclassification of the true negative pixels and is expected to show robust performance. During the training process, data augmentations such as horizontal flip, vertical flips, shifting, rotation, contrast transformations, and grid distortion were applied to all the slices, regardless of WMH presence, to foster the development of a noise-robust model. The second U-Net model used EfficientNet-B5 as the encoder, which was trained in the same manner as the first U-Net. The output of the main CNN model was a threshold of 0.5, and the predicted binary WMH mask was obtained. The number of ensembles was set to two to obtain a balance between processing time and improved accuracy. Ensembles comprising three or more models increased the processing time by approximately 1.5 times or more, although the enhancement in performance was marginal. Pseudo-labeling was employed to effectively utilize the unlabeled data in the training of the actual model. First, a model for pseudo-labeling was trained using the PR for training only. The model was then used to predict the WMH regions in 885 unannotated participants. These predicted WMH masks can be regarded as pseudo-WMH annotations (pseudo-labeling dataset). Finally, a new model was trained using both the PR and pseudo-labeling dataset. Evaluation metrics To evaluate WMH segmentation performance, the voxel-wise DSC was used in this study. The DSC is a measure of the agreement between predicted and ground-truth regions and is defined as DSC = 2 × TP / (2 × TP + FP + FN), where TP, FP, and FN are the numbers of true positive, FP, and FN voxels, respectively. Moreover, recall and precision were calculated using the following formulas: recall = TP / (TP + FN) and precision = TP / (TP + FP). Processing time measurement To evaluate the applicability of the model in clinical practice with limited resources, processing time was measured without using a graphics processing unit (GPU). The average of 10 measurements was computed for thick-slice MRI images. To facilitate comparison with a previous study [ 32 ], the measured value was corrected for differences in the number of slices and central processing unit (CPU) clock frequencies as follows. In the prior study, time measurements were obtained using MRI with 192 slices on a computer with a 3.5-GHz CPU, whereas our study used 22 slices on a computer without a GPU, equipped with an Intel Core i5-10500T processor (Intel Corp., Santa Clara, CA, USA) running at 2.30 GHz with 16 GB of memory. Assuming that processing time decreases proportionally with the CPU clock frequency, the corrected processing time, intended for comparison with the results of the previous study, was calculated as follows: corrected processing time = measured time (min) × (192/22) × (2.3/3.5). Abbreviations 2D: two-dimensional, 3D: three-dimensional, AI: artificial intelligence, CNN: convolutional neural network, CPU: central processing unit, DSC: Dice similarity coefficient, DSWMH: deep and subcortical white matter hyperintensity, EPVS: enlarged perivascular space, FLAIR: fluid-attenuated inversion recovery, FN: false negative, FP: false positive, GPU: graphics processing unit, MRI: magnetic resonance imaging, PR: Private Dataset, PVH: periventricular hyperintensity, T1WI: T1-weighted image, T2WI: T2-weighted image, TP: true positive, WMH: white matter hyperintensity, WMHC: WMH Segmentation Challenge Declarations Acknowledgments We thank all the radiologists and neurosurgeons for providing the diagnoses used in this study. Author contributions All authors have made substantial contributions to the intellectual content of the paper, contributed to data interpretation, approved the final manuscript, and agreed to its submission to this journal. M.K. contributed to the study design and concept; funding acquisition; research conduction; data collection, curation, management, and analysis; quality control; statistical analysis; and manuscript drafting. F.I. contributed to the study design and concept; funding acquisition; data collection and analysis; quality control; statistical analysis; and manuscript revision. S.N. and S.K. helped conceive and oversee the study and contributed to the revision of the manuscript. D.I., H.K., T.H., and Y.M. contributed to data collection, curation, and analysis. R.S., S.M., and T.K. helped with data collection and analysis and the revision of the manuscript. Y.S. helped conceive and oversee the study and assisted in data collection. N.H. assisted in the planning and supervision of this study, as well as in the revision of the manuscript. Data Availability Anonymized data from the present study may be shared by the corresponding author upon request from a qualified researcher and upon permission from the institutional review board. Funding This work was supported by the Japan Society for the Promotion of Science, Grant-in-Aid for Scientific Research, (C)23K08521. Competing Interests Shinji Nakazawa, Ryo Sato, and Shiyuki Maeyama are employees of LPIXEL Inc. Yuki Shimahara owns stock in the company. Taiki Kaneko is currently an employee of Aile Home Clinic, Niigata, Japan. Masashi Kuwabara, Fusao Ikawa, Saori Koshino, Daizo Ishii, Hiroshi Kondo, Takeshi Hara, Yuyo Maeda, and Nobutaka Horie declare no potential conflict of interest. References Wardlaw, J. M. et al. Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration. Lancet Neurol. 12, 822–838 (2013). van Dijk, E. J. et al. Progression of cerebral small vessel disease in relation to risk factors and cognitive consequences: Rotterdam Scan study. Stroke 39, 2712–2719 (2008). Mosley, T. H. et al. Cerebral MRI findings and cognitive functioning: the Atherosclerosis Risk in Communities study. Neurology 64, 2056–2062 (2005). Doddy, R. S., Massman, P. J., Mawad, M. & Nance, M. Cognitive consequences of subcortical magnetic resonance imaging changes in Alzheimer's disease: comparison to small vessel ischemic vascular dementia. Neuropsychiatry Neuropsychol. Behav. Neurol. 11, 191–199 (1998). O'Brien, J. et al. Severe deep white matter lesions and outcome in elderly patients with major depressive disorder: follow up study. BMJ 317, 982–984 (1998). Ter Telgte, A. et al. Cerebral small vessel disease: from a focal to a global perspective. Nat. Rev. Neurol. 14, 387–398 (2018). Prins, N. D. & Scheltens, P. White matter hyperintensities, cognitive impairment and dementia: an update. Nat. Rev. Neurol. 11, 157–165 (2015). Simoni, M. et al. Age- and sex-specific rates of leukoaraiosis in TIA and stroke patients: population-based study. Neurology 79, 1215–1222 (2012). Yamasaki, T. et al. Prevalence and risk factors for brain white matter changes in young and middle-aged participants with Brain Dock (brain screening): a registry database study and literature review. Aging (Albany NY) 13, 9496–9509 (2021). Breteler, M. M. et al. Cognitive correlates of ventricular enlargement and cerebral white matter lesions on magnetic resonance imaging. The Rotterdam Study. Stroke 25, 1109–1115 (1994). New Guidelines Development Committee for Brain Dock. [The Guideline for Brain Dock 2019]: Kyobunsha; 2019. Morita, A. Value of Brain Dock (Brain Screening) System in Japan. World Neurosurg. 127, 502 (2019). Saito, I. [The Guideline for Brain Dock 2003]. Nihon Rinsho 64 Suppl 7, 297–302 (2006). Zhu, W. et al. Automatic segmentation of white matter hyperintensities in routine clinical brain MRI by 2D VB-Net: a large-scale study. Front. Aging Neurosci. 14, 915009 (2022). Joo, L. et al . Diagnostic performance of deep learning-based automatic white matter hyperintensity segmentation for classification of the Fazekas scale and differentiation of subcortical vascular dementia. PLoS One 17, e0274562 (2022). Zijdenbos, A. P., Forghani, R. & Evans, A. C. Automatic "pipeline" analysis of 3-D MRI data for clinical trials: application to multiple sclerosis. IEEE Trans. Med. Imaging 21, 1280–1291 (2002). Grimaud, J. et al. Quantification of MRI lesion load in multiple sclerosis: a comparison of three computer-assisted techniques. Magn. Reson. Imaging 14, 495–505 (1996). Røvang M. S., et al. Segmenting white matter hyperintensities on isotropic three-dimensional fluid attenuated inversion recovery magnetic resonance images: assessing deep learning tools on a Norwegian imaging database. PLoS One. 18, e0285683 (2023). Ding, Y. et al. Using deep convolutional neural networks for neonatal brain image segmentation. Front. Neurosci. 14, 207 (2020). Ding, T. et al. An improved algorithm of white matter hyperintensity detection in elderly adults. Neuroimage Clin. 25, 102151 (2020). Park, G., Hong, J., Duffy, B. A., Lee, J. M. & Kim, H. White matter hyperintensities segmentation using the ensemble U-Net with multi-scale highlighting foregrounds. Neuroimage 237, 118140 (2021). Liu, L., Kurgan, L., Wu, F. X. & Wang, J. Attention convolutional neural network for accurate segmentation and quantification of lesions in ischemic stroke disease. Med. Image Anal. 65, 101791 (2020). Le M., et al . FLAIR 2 improves LesionTOADS automatic segmentation of multiple sclerosis lesions in non-homogenized, multi-center, 2D clinical magnetic resonance images. Neuroimage Clin. 23, 101918 (2019). Heinen R, et al . Performance of five automated white matter hyperintensity segmentation methods in a multicenter dataset. Sci. Rep. 9, 16742 (2019). Zhang Y, et al . A deep learning algorithm for white matter hyperintensity lesion detection and segmentation. Neuroradiology 64, 727–734 (2022). Park, B. Y. et al. DEWS (DEep White matter hyperintensity Segmentation framework): a fully automated pipeline for detecting small deep white matter hyperintensities in migraineurs. Neuroimage Clin. 18, 638–647 (2018). Moeskops, P. et al. Evaluation of a deep learning approach for the segmentation of brain tissues and white matter hyperintensities of presumed vascular origin in MRI. Neuroimage Clin. 17, 251–262 (2018). Gibson, E., Gao, F., Black, S. E. & Lobaugh, N. J. Automatic segmentation of white matter hyperintensities in the elderly using FLAIR images at 3T. J. Magn. Reson. Imaging 31, 1311–1322 (2010). Kuijf, H. J. et al. Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH Segmentation Challenge. IEEE Trans. Med. Imaging 38, 2556–2568 (2019). Li, H. et al. Fully convolutional network ensembles for white matter hyperintensities segmentation in MR images. Neuroimage 183, 650–665 (2018). Rachmadi, M. F., Valdés-Hernández, M. D. C., Agan, M. L. F., Di Perri, C. & Komura, T. Segmentation of white matter hyperintensities using convolutional neural networks with global spatial information in routine clinical brain MRI with none or mild vascular pathology. Comput. Med. Imaging Graph 66, 28–43 (2018). Tran, P. et al. Automatic segmentation of white matter hyperintensities: validation and comparison with state-of-the-art methods on both Multiple Sclerosis and elderly subjects. Neuroimage Clin. 33, 102940 (2022). Fiford, C. M. et al. Automated white matter hyperintensity segmentation using Bayesian model selection: assessment and correlations with cognitive change. Neuroinformatics 18, 429–449 (2020). Wu, J., Zhang, Y. & Tang, X. Simultaneous tissue classification and lateral ventricle segmentation via a 2D U-net driven by a 3D fully convolutional neural network. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2019, 5928–5931 (2019). Wu, D. et al. Multi-atlas based detection and localization (MADL) for location-dependent quantification of white matter hyperintensities. Neuroimage Clin. 22, 101772 (2019). Ling, Y., Jouvent, E., Cousyn, L., Chabriat, H. & De Guio, F. Validation and optimization of BIANCA for the segmentation of extensive white matter hyperintensities. Neuroinformatics 16, 269–281 (2018). Manjón, J. V. et al. MRI white matter lesion segmentation using an ensemble of neural networks and overcomplete patch-based voting. Comput. Med. Imaging Graph 69, 43–51 (2018). Abhishek, K. & Hamarneh, G. Matthews correlation coefficient loss for deep convolutional networks: application to skin lesion segmentation. 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). IEEE Xplore® https://ieeexplore.ieee.org/document/9433782/authors#authors , (2021). Additional Declarations No competing interests reported. Supplementary Files srsupplementarytables12024.1.1.docx Cite Share Download PDF Status: Published Journal Publication published 02 May, 2024 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 04 Mar, 2024 Reviews received at journal 28 Feb, 2024 Reviewers agreed at journal 19 Feb, 2024 Reviews received at journal 12 Feb, 2024 Reviewers agreed at journal 15 Jan, 2024 Reviewers invited by journal 14 Jan, 2024 Editor assigned by journal 11 Jan, 2024 Editor invited by journal 10 Jan, 2024 Submission checks completed at journal 10 Jan, 2024 First submitted to journal 04 Jan, 2024 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-3833822","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":266507123,"identity":"70e67bde-6817-41ca-8d4d-77be8e139a37","order_by":0,"name":"Masashi Kuwabara","email":"","orcid":"","institution":"Hiroshima University","correspondingAuthor":false,"prefix":"","firstName":"Masashi","middleName":"","lastName":"Kuwabara","suffix":""},{"id":266507124,"identity":"953ea98d-f0f7-4298-9eb8-47e983456d4e","order_by":1,"name":"Fusao Ikawa","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4klEQVRIie3QPQrCMByH4Z8U0iXgKhbsFSIZHbyKQXCKUnB1KAh2c+7mFZycC4G69AAVXUToBQQXv6lYBCHaTSTv9B/y5AswmX4we0yeoxUB7DEyHaGqIKTzJYkKQrULX4hNxJ6OMHBDeain3gbVIMLQ0xGLKIfGGLK0v3Aky1BLOuChhrQt23ckuYp5LScKSAFOtafY46O8QMxCmeXE/UxI7PQnEH4qSU7YF6TXOk8h5knGW3dCm4nwtW+h1ZivwsP9YkF3u5Yn1WgsVcx1P/a+B1CZ8DIiz9qVJiaTyfTP3QCyzkOr6ui1pgAAAABJRU5ErkJggg==","orcid":"","institution":"Shimane Prefectural Central Hospital","correspondingAuthor":true,"prefix":"","firstName":"Fusao","middleName":"","lastName":"Ikawa","suffix":""},{"id":266507125,"identity":"09547e12-585f-425f-bdae-716a23ac9f48","order_by":2,"name":"Shinji Nakazawa","email":"","orcid":"","institution":"LPIXEL Inc","correspondingAuthor":false,"prefix":"","firstName":"Shinji","middleName":"","lastName":"Nakazawa","suffix":""},{"id":266507126,"identity":"37eb095a-4cc8-4cca-b2f3-9212b7af64ad","order_by":3,"name":"Saori Koshino","email":"","orcid":"","institution":"The University of Tokyo Hospital","correspondingAuthor":false,"prefix":"","firstName":"Saori","middleName":"","lastName":"Koshino","suffix":""},{"id":266507127,"identity":"3bb73349-1c62-4b1d-802e-46d377d8d885","order_by":4,"name":"Daizo Ishii","email":"","orcid":"","institution":"Hiroshima University","correspondingAuthor":false,"prefix":"","firstName":"Daizo","middleName":"","lastName":"Ishii","suffix":""},{"id":266507128,"identity":"26aa2cf5-9d22-4728-9108-02a0e3154eed","order_by":5,"name":"Hiroshi Kondo","email":"","orcid":"","institution":"Hiroshima University","correspondingAuthor":false,"prefix":"","firstName":"Hiroshi","middleName":"","lastName":"Kondo","suffix":""},{"id":266507129,"identity":"9f4f2633-0b67-4d8e-b9b2-cbdbc5c3f324","order_by":6,"name":"Takeshi Hara","email":"","orcid":"","institution":"Hiroshima University","correspondingAuthor":false,"prefix":"","firstName":"Takeshi","middleName":"","lastName":"Hara","suffix":""},{"id":266507130,"identity":"20d66126-ce85-47e1-aabb-65e766cd2058","order_by":7,"name":"Yuyo Maeda","email":"","orcid":"","institution":"Hiroshima University","correspondingAuthor":false,"prefix":"","firstName":"Yuyo","middleName":"","lastName":"Maeda","suffix":""},{"id":266507131,"identity":"0126219a-2b17-49eb-b9bf-b0c8afc935da","order_by":8,"name":"Ryo Sato","email":"","orcid":"","institution":"LPIXEL Inc","correspondingAuthor":false,"prefix":"","firstName":"Ryo","middleName":"","lastName":"Sato","suffix":""},{"id":266507132,"identity":"51ae62aa-b8a9-4b71-ab15-f73de769132b","order_by":9,"name":"Taiki Kaneko","email":"","orcid":"","institution":"LPIXEL Inc","correspondingAuthor":false,"prefix":"","firstName":"Taiki","middleName":"","lastName":"Kaneko","suffix":""},{"id":266507133,"identity":"4bf2b27b-d225-4415-8e2b-070305f19690","order_by":10,"name":"Shiyuki Maeyama","email":"","orcid":"","institution":"LPIXEL Inc","correspondingAuthor":false,"prefix":"","firstName":"Shiyuki","middleName":"","lastName":"Maeyama","suffix":""},{"id":266507134,"identity":"13ed29c4-9284-4962-96b7-c4cf6b0bdb96","order_by":11,"name":"Yuki Shimahara","email":"","orcid":"","institution":"LPIXEL Inc","correspondingAuthor":false,"prefix":"","firstName":"Yuki","middleName":"","lastName":"Shimahara","suffix":""},{"id":266507135,"identity":"bbf9de28-36a1-4021-9714-79b1d9e228e8","order_by":12,"name":"Nobutaka Horie","email":"","orcid":"","institution":"Hiroshima University","correspondingAuthor":false,"prefix":"","firstName":"Nobutaka","middleName":"","lastName":"Horie","suffix":""}],"badges":[],"createdAt":"2024-01-04 06:47:33","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3833822/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3833822/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-024-60789-x","type":"published","date":"2024-05-02T19:57:53+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":49547447,"identity":"04e542dc-13ac-46d0-a502-f6a779b977a1","added_by":"auto","created_at":"2024-01-12 19:21:10","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":159511,"visible":true,"origin":"","legend":"\u003cp\u003eHistograms showing the WMH volume distributions throughout the annotated Private Dataset (n=207) and WMHC Dataset (n=170).\u003cstrong\u003e \u003c/strong\u003eThe WMH volume histograms of the Private Dataset (\u003cstrong\u003eA\u003c/strong\u003e) and WMHC Dataset (\u003cstrong\u003eB\u003c/strong\u003e) are shown. The x-axis represents the total WMH volume in each participant, and the y-axis represents the counts. WMH: white matter hyperintensity, WMHC: white matter hyperintensity segmentation challenge.\u003c/p\u003e","description":"","filename":"srwmhfigure1.2024.1.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3833822/v1/7f8b75f153d9ee96896c60b4.jpg"},{"id":49547446,"identity":"cff6ba60-ba2e-4996-a108-5a78b12540d1","added_by":"auto","created_at":"2024-01-12 19:21:10","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":361814,"visible":true,"origin":"","legend":"\u003cp\u003eSegmentation results for Private Dataset and WMHC Dataset models used with the test dataset. The columns from left to right show the original FLAIR image, ground-truth WMH annotations, and segmentation results for the PR, PR+WMHC, and WMHC models. The first row shows a TP case in which all models correctly detected WMHs, the second shows FN regions in the WMHC results, and the third row shows small FN regions in the PR and PR+WMHC results. The last row shows the difference in FP regions between PR and PR+WMHC results. FLAIR: fluid-attenuated inversion recovery, FN: false negative, FP: false positive, PR: Private Dataset, TP: true positive, WMH: white matter hyperintensity, WMHC: WMH Segmentation Challenge.\u003c/p\u003e","description":"","filename":"srwmhfigure2.2024.1.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3833822/v1/c3e276fc04033fa972efd987.jpg"},{"id":49547444,"identity":"81814dee-5229-481d-ad41-5c62ee2d1abd","added_by":"auto","created_at":"2024-01-12 19:21:10","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2456684,"visible":true,"origin":"","legend":"\u003cp\u003eSegmentation results of the PR model for cases with the lowest and highest DSC values from the Private Dataset. (\u003cstrong\u003eA\u003c/strong\u003e) The case with the lowest DSC value of 0.158; (\u003cstrong\u003eB\u003c/strong\u003e) the case with the highest DSC value of 0.901. The columns titled “Original FLAIR,” “WMH annotation,” and “PR” display the original FLAIR images, the ground-truth WMH annotations and the segmentation results of the PR model are shown in green and red, respectively. DSC: Dice similarity coefficient, FLAIR: fluid-attenuated inversion recovery, WMH: white matter hyperintensity, PR: Private Dataset\u003c/p\u003e","description":"","filename":"srwmhfigure3.2024.1.1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3833822/v1/88d21a056766df032bbf4989.jpg"},{"id":56043031,"identity":"c0324dab-1ff5-434a-acac-c730175ef064","added_by":"auto","created_at":"2024-05-07 20:09:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1047707,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3833822/v1/9529f98c-6dfe-4215-8f07-9893377056f8.pdf"},{"id":49548187,"identity":"2a2722f5-f9ff-4ec8-88e3-10f725292ad0","added_by":"auto","created_at":"2024-01-12 19:29:10","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":23466,"visible":true,"origin":"","legend":"","description":"","filename":"srsupplementarytables12024.1.1.docx","url":"https://assets-eu.researchsquare.com/files/rs-3833822/v1/ef548021cc98fdfcd00a2573.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial intelligence for volumetric measurement of cerebral white matter hyperintensities on thick-slice fluid-attenuated inversion recovery (FLAIR) magnetic resonance images from multiple centers","fulltext":[{"header":"Introduction","content":"\u003cp\u003eWhite matter hyperintensities (WMHs) are common neuroimaging findings characterized by high signals either as periventricular hyperintensities (PVHs) or deep subcortical WMHs (DSWMHs) in fluid-attenuated inversion recovery (FLAIR) sequences of head magnetic resonance imaging (MRI) [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Many reports have shown associations of the WMH degree with ischemic stroke, depression, dementia, and emotional disorders [\u003cspan additionalcitationids=\"CR3 CR4 CR5 CR6 CR7\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Although the WMH prevalence is higher in older individuals, studies of healthy younger people have reported that the WMH degree is also correlated with the execution of speed-demanding functions, such as word recall, that is subfrontal cortical functions [\u003cspan additionalcitationids=\"CR7 CR8 CR9\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBrain Dock is a widely implemented brain checkup system that uses MRI and magnetic resonance angiography to detect asymptomatic brain diseases, such as unruptured cerebral aneurysms and WMHs at an early stage, thereby preventing stroke and dementia in healthy individuals in Japan [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. During a Brain Dock, the evaluation of WMHs is required, using a standard database with the aim of accumulating data on the diagnosis and early detection of disease risk [\u003cspan additionalcitationids=\"CR12\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. To date, manual interpretation by neuroradiologists has been considered an effective method for quantitatively assessing WMHs [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. However, this approach is time-consuming, labor-intensive, and has the disadvantage of being unsuitable for clinical application or large-scale studies because of the large interrater variability ranging from 10\u0026ndash;68% [\u003cspan additionalcitationids=\"CR15 CR16\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecently, artificial intelligence (AI) has been used for automated reading and volumetric measurements. Although many publications have reported the use of AI to assess brain atrophy on T1-weighted images (T1WIs), reports on AI algorithms for automated WMH segmentation using only FLAIR images are limited [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The size, number, shape, and location of WMHs are heterogeneous and age-dependent, making them difficult to evaluate using AI. Moreover, WMHs are present in a variety of diseases, requiring background information, such as patient history and disease course, and multiple MRI sequences for differentiation [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Only a few of the reported AI algorithms are applicable in clinical practice because many are limited to specific diseases, have small sample sizes, evaluate multiple sequences, or require a thin slice thickness as the imaging conditions [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. To ensure applicability in clinical practice, accurate results must be provided consistently under imaging protocols for various WMHs, often requiring multicenter collaborative studies [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Only a few studies to date have validated the AI algorithm for automatic WMH segmentation by FLAIR alone in a multicenter setting because of the difficulty [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. If AI algorithms for automated WMH segmentation using only FLAIR images were developed and automatically classified in Brain Dock, the scientific contribution to preventive medicine in healthy people would be immeasurable. The initial purpose of this study was to develop a new AI software that can automatically extract and measure WMH volumes on head MRI using only thick-slice FLAIR images from multiple centers. In the future, we plan to develop an AI that can automatically classify WMHs.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eAccording to the five neuroradiologists participating in this study, the average WMH volume was 18.1\u0026thinsp;\u0026plusmn;\u0026thinsp;22.1 mL, with a minimum of 0.0 mL and a maximum of 109.5 mL. The distribution of WMH volumes in the Private Dataset (PR) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea) closely resembled that in the WMH Segmentation Challenge (WMHC) Dataset (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb). Given the potential for biased evaluation data to negatively impact generalization performance on unseen data, the PR was expected to evaluate generalization performance with similar accuracy to that of the WMHC dataset. Moreover, the 885 unannotated participants were utilized for pseudo-labeling to enhance model performance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eQuantitative analysis\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the evaluation results of WMH segmentation based on the Dice similarity coefficient (DSC), Recall, precision, and modified Hausdorff distance (95th percentile of the Hausdorff distance in mm) were compared. The PR model demonstrated high performance with a DSC of 0.8 or higher on the PR, whereas its performance on the WMHC Dataset was not as high. The PR\u0026thinsp;+\u0026thinsp;WMHC model, which included the WMHC Dataset for training, showed improved performance and the ability to handle both thick- and thin-slice MRI datasets simultaneously. However, its performance on the PR improved only marginally, with a DSC increase of only 0.002. This modest degree of improvement might be attributed to domain differences between these datasets, suggesting that the thin-slice dataset does not significantly enhance performance on the thick-slice dataset. Additionally, the model trained only on the WMHC Dataset exhibited overfitting to the test dataset.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAssessment of WMH segmentation by the U-Net ensemble models\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eModel name\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTraining dataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c6\" namest=\"c3\"\u003e \u003cp\u003ePrivate Dataset for the test dataset (n\u0026thinsp;=\u0026thinsp;69)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c10\" namest=\"c7\"\u003e \u003cp\u003eWMHC Dataset for the test dataset (n\u0026thinsp;=\u0026thinsp;110)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDSC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eH95\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eH95\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrivate Dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.820\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.788\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.855\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e14.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.742\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.653\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e22.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePR\u0026thinsp;+\u0026thinsp;WMHC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrivate and WMHC Datasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.822\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.789\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.858\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e14.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.877\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.781\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e7.97\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWMHC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWMHC Dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.531\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.931\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e16.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.833\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.874\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e5.61\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"10\"\u003eDSC: Dice similarity coefficient, H95: modified Hausdorff distance (95th percentile [mm]), PR: Private Dataset, WMH: white matter hyperintensity, WMHC: white matter hyperintensity segmentation challenge.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eQualitative analysis\u003c/h2\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the segmentation results of the PR and WMHC Dataset models used with the test dataset. The top row presents a participant with a relatively small WMH volume. All models successfully detected regions that were similar to the ground truth. In a participant with extensive WMHs, as in the second row, both the PR and PR\u0026thinsp;+\u0026thinsp;WMHC models accurately detected the WMH regions, whereas the WMHC model exhibited false negatives (FNs), leading to a lower DSC. However, in a participant with small punctate WMHs, as observed in the right hemisphere in the third row, the WMHC model demonstrated better results compared with the other two models. The differences between the PR and the PR\u0026thinsp;+\u0026thinsp;WMHC models are shown in the last row. False positives (FPs), which occurred in the PR model, were reduced in the PR\u0026thinsp;+\u0026thinsp;WMHC model. For all other participants, the difference between these two models was minimal, which resulted in a slight DSC difference of 0.002.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA tendency for FPs and FNs was observed in the PR model in areas with small punctate WMHs, near the lateral ventricles, near the septum pellucidum, lacunar infarcts, or where the boundary between WMHs and other structures was ambiguous. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, row 3, the model missed the small punctate WMHs in the left hemisphere. Conversely, in the example from Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, row 4, punctate FPs can be observed in the left hemisphere. The figure also shows FPs near the surface of the lateral ventricle body, a region where the boundary between a WMH and other structures could be indistinct. While the examples in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, rows 1 and 2, are generally accurate, an unintended \u0026ldquo;hole\u0026rdquo; appeared in the right hemisphere of Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, row 2. Furthermore, the recall of the PR model was lower than its precision, indicating that the model tended to have an under-detection bias on the PR.\u003c/p\u003e \u003cp\u003eTo better understand variations on a case-by-case basis, the DSC for each case was calculated using the PR model results of the PR. The DSC values ranged from a minimum of 0.158 to a maximum of 0.901. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea and \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb show the cases with the lowest and highest DSCs, respectively. The case with the lowest DSC, at 0.158, missed small punctate WMHs, and the missed area accounted for a large portion of the total WMH area. Conversely, the case with the highest DSC (0.901) showed a larger total WMH area. Although some FNs or FPs were observed, their proportions to the total count were limited, resulting in a high DSC.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eProcessing time analysis\u003c/h2\u003e \u003cp\u003eProcessing time measurements were conducted for thick-slice MRI images, revealing an average processing time of 18.5 s per case. When corrected for comparison with a previous study, the equivalent value was approximately 1.8 min.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we evaluated the performance of an AI using automatically extracted volumetric measurement of WMH lesions on head MRI images on thick-slice FLAIR images. The results showed that the DSC in our study reached 0.820, which is comparable to human reading accuracy.\u003c/p\u003e \u003cp\u003eIn recent years, although various algorithms for AI-based automated volume measurements of WMHs have been reported in the literature, very few are applicable in clinical practice [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In previous studies on AI and volumetric measurements of WMHs, many were limited to specific diseases, such as Alzheimer\u0026rsquo;s disease or multiple sclerosis, and had relatively small sample sizes of approximately 100 patients [\u003cspan additionalcitationids=\"CR27\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] and an algorithms that used three-dimensional (3D) networks with thin-slice image data with 1\u0026ndash;3 mm slice thicknesses [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Thin-slice MRI has only been widely applied in a few developed countries and is currently not commonly used in clinical practice [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. A qualitative evaluation with subjective judgments should be objectively compared with a quantitative evaluation based on automatic volumetric measurement using AI.\u003c/p\u003e \u003cp\u003eWMHs result primarily from aging processes, such as demyelination and axonal loss, both of which occur as a result of small vessel disease of the brain [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Thus, WMHs are frequently observed in older asymptomatic patients, although they can also occur in a variety of other diseases, including demyelinating diseases such as multiple sclerosis; genetic diseases such as cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy; infectious diseases such as human immunodeficiency virus encephalopathy; psychiatric disorders such as depressive disorders and schizophrenia; and various other diseases such as autoimmune diseases, neoplastic diseases, intoxication, and hypoxic encephalopathies [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Therefore, volumetric detection of WMHs in younger people is an important aspect of brain checkups, and lifestyle improvements are needed in examinees with increased WMH volume over time.\u003c/p\u003e \u003cp\u003eThis study proposes a new AI software that solves these problems, and its major feature is 5-mm thick-slice FLAIR images. Moreover, this study evaluated both the PR and the WMHC Dataset as thick (5 mm) and thin (1\u0026ndash;3 mm) slices, respectively, with a sample size of over 200 participants. The results showed that the WMH segmentation accuracy based on two-dimensional [2D] FLAIR images with 5 mm slices, which is widely used in general hospitals, was comparable to that of previous studies. This suggests that the proposed AI algorithm can be applied in clinical practice.\u003c/p\u003e \u003cp\u003eWe only identified 39 publications that analyzed the automatic segmentation of WMHs on head MRI using AI similar to the present study. Of these, we reviewed 17 publications since 2018, including the present study (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The sample size varied from 20 to 1,092 overall, with the sample size of this study being the largest. Multiple MRI sequences analyzing both T1WI and FLAIR images were identified in 12 studies [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan additionalcitationids=\"CR33 CR34 CR35 CR36\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], and only 1 study in addition to the present study evaluated only FLAIR images. Regarding the MRI field strength, nearly all studies, including ours, used 3T MRI. Regarding the segmentation accuracy, the DSC values ranged from 0.41 to 0.83, showing a wide range of accuracy. The DSC of the AI model in this study was 0.820, indicating higher accuracy compared with that of previous studies overall. Therefore, the model developed in the present study demonstrably has a better performance than that of existing AI models.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eReview of recent studies on the automatic segmentation of cerebral white matter hyperintensities in head MRI using AI \u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAuthor / Year\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eSample size\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eBreakdown of samples\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eType of segmentation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAverage spatial agreement with reference segmentation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSequence\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eField strength (T)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eManj\u0026oacute;n JV / 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKnight J / 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.41\u0026ndash;0.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eQin C / 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLing Y / 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e156\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.76\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePark BY / 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e148\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.65\u0026thinsp;\u0026minus;\u0026thinsp;0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtlason HE / 2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e170\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain: 60, test: 110\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, T2WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eUnsupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.53\u0026thinsp;\u0026minus;\u0026thinsp;0.67, TPR: 0.25\u0026ndash;0.40\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSundaresan V / 2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e133\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, T2WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.77, TPR: 0.73\u0026ndash;0.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWu D / 2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e135\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain: 15, test: 120\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.62, FPR: 0.35, FNR: 0.37\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWu J / 2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDing T / 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain: 15, test: 5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.78, TPR: 0.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFiford CM / 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain: 20, test: 60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eUnsupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHong J / 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e148\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTPR: 0.87, FDR: 0.10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiu L / 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etest: 60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.83\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRachmadi MF / 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etest: 60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, T2WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eUnsupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.47\u0026thinsp;\u0026minus;\u0026thinsp;0.56, TPR: 0.47\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTran P / 2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZhu W / 2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,045\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecross-validation: 849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eT1WI, FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.792\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePresent study / 2023\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,092\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003etrain: 138\u0026thinsp;+\u0026thinsp;885, test: 69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.5, 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSupervised\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDSC: 0.820\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \n\u003cp\u003eDSC: dice score coefficient, TPR: true positive rate, FPR: false positive rate, FNR: false negative rate, FDR: false discovery rate, FLAIR: fluid-attenuated inversion recovery, T1WI: T1-weighted imaging, T2WI: T2-weighted imaging, T: tesla, N/A: not applicable, MRI: magnetic resonance imaging, AI: artificial intelligence\u003c/p\u003e\n\u003cp\u003eThe unique feature of our study is that we have developed an AI that performs better than previous AI models using \u0026ldquo;thick slices\u0026rdquo; and \u0026ldquo;FLAIR images only\u0026rdquo; in accordance with the standards of the Japan Brain Dock Society. Notably, the performance of this AI model was tested on the largest sample size in a study of this type of more than 1,000 cases, using data from multiple institutions. In addition, the processing time when performing segmentation was as fast as 18.5 s per case for our AI model. When corrected for comparison with a previous study, the equivalent value was approximately 1.8 min. This is faster than the approximate processing time of 5 min using a supervised deep learning-based method reported in a previous study [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. Therefore, this AI algorithm seems more appropriate for real-world clinical adaptation. MRI imaging with thick slices and FLAIR images only are simple methods for brain checkups, and thus has an advantage in that MRI examinations are easily accessible and versatile for patients. Our findings can also be used as a basis for PVH and DSWMH grading as proposed by the Japan Brain Dock Society. The ultimate goal of our AI model is to classify PVH and DSWMH grading to screen normal participants at risk of early-onset dementia to prevent or delay the onset of dementia through lifestyle guidance.\u003c/p\u003e \u003cp\u003eThis study had several limitations. First, approximately 10 physicians diagnosed the images for annotation, possibly leading to individual differences in physician judgment, resulting in diagnostic bias. However, all physicians in this study were board certified radiologists from the Japan Radiological Society or neuroradiologists with more than 7 years of experience and board-certified neurosurgeons from the Japan Neurosurgical Society; therefore, the diagnostic results should be of a uniformly high standard. Second, because only high intensity areas on FLAIR images were evaluated in this study, lacunar infarction might not have been clearly distinguished because only two lacunar infarctions were included in the training dataset. We would like to develop another AI to distinguish between WMHs and lacunar infarctions after acquiring a lacunar infarctions dataset in the near future. Third, a detailed analysis differentiating between PVH and DSWMH within WMHs was not conducted in this study; however, our research is currently ongoing.\u003c/p\u003e \u003cp\u003eIn conclusion, the automatic WMH segmentation model based on a U-Net ensemble trained on a thick-slice FLAIR MRI dataset showed a DSC of 0.820 and a promising AI algorithm. This model may be applicable in clinical practice for brain checkups.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eEthics approval and informed consent\u003c/h2\u003e \u003cp\u003e This study was approved by the Institutional Review Board of Hiroshima University (approval number: E2022-0262). All study protocols were developed according to the guidelines of Hiroshima University Hospital. Individual data were anonymized and collected during routine Brain Dock examinations; thus, all participants provided informed consent based on an opt-out method. Moreover, in this study, we were provided with anonymized data; therefore, the authors did not have access to personally identifiable participant information during or after data collection.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy design\u003c/h3\u003e\n\u003cp\u003eIn this study, a new automatic WMH segmentation model was developed based on existing commercial software (EIRL Brain Metry version 1.13.0, LPIXEL Inc., Tokyo, Japan). This software employs U-Net for segmentation. Although the core usage of U-Net remains consistent, we replaced the backbone network with a new model and also explored the model ensembles.\u003c/p\u003e \u003cp\u003eThe WMH segmentation model of the software was updated using a newly collected dataset, the PR. This new model was named \u0026ldquo;PR\u0026rdquo; as it was trained using the PR. The model performance was also evaluated using the publicly available WMHC Dataset [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. The WMHC Dataset has thinner slices than those in the PR.\u003c/p\u003e \u003cp\u003eFor validation, two additional models were trained using the same training procedure. One was a model trained using both the PR and the WMHC Dataset (PR\u0026thinsp;+\u0026thinsp;WMHC model) to evaluate the effect of adding an MRI dataset with thin slices to the PR with thick slices. The other was a model trained using only the WMHC Dataset (the WMHC model) to evaluate the performance of a model trained using only a thin-slice MRI dataset.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePrivate Dataset\u003c/h2\u003e \u003cp\u003eFor this dataset, 1,092 FLAIR MRI images were collected from three clinics in Japan. To minimize potential biases related to the WMH severity, data was gathered in a manner that ensured a balanced distribution of PVH/DSWMH grades (0\u0026ndash;4) as determined by each clinic. The WMH annotation process was restricted to 207 randomly selected participants because of the annotation costs. Nearly all the annotated WMHs were age-related, and 2 of these 207 participants had lacunar infarcts. The slice thickness of the annotated data ranged from 5\u0026ndash;6 mm, with an average of 5.37 mm for the entire dataset. Detailed spatial characteristics, including slice thickness, are presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The data was acquired using nine different types of scanners, and the detailed MRI parameters are listed in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCharacteristics of the Private Dataset comprising 207 annotated participants\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026times;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVendor\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eScanner type ID\u003csup\u003ea\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eScanner\u003c/p\u003e \u003cp\u003emodel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMagnetic field strength (T)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eVoxel size (mm\u003csup\u003e3\u003c/sup\u003e)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGE Healthcare\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDISCOVERY MR750w\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e6.0 \u0026times; 0.43 \u0026times; 0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Chicago, IL, USA)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOptima MR450w\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.469 \u0026times; 0.469\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSIGNA Architect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.43 \u0026times; 0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSIGNA EXCITE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e6.0 \u0026times; 0.469 \u0026times; 0.469\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSigna HDxt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.469 \u0026times; 0.469\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSigna HDxt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e6.0 \u0026times; 0.43 \u0026times; 0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSigna HDxt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e6.0 \u0026times; 0.469 \u0026times; 0.469\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSigna HDxt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e6.0 \u0026times; 0.859 \u0026times; 0.859\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSigna HDxt\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.43 \u0026times; 0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePhilips\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIngenia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.479 \u0026times; 0.479\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Amsterdam, the Netherlands)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIngenia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.448 \u0026times; 0.448\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIngenia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.449 \u0026times; 0.449\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSiemens Healthineers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAvanto\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.898 \u0026times; 0.898\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Erlangen, Germany)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSymphony\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e5.0 \u0026times; 0.719 \u0026times; 0.719\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e\u003csup\u003ea\u003c/sup\u003eScanner type ID: a unique ID for each combination of scanner and magnetic field strength\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFive neuroradiologists shared the annotation efforts of WMHs for the 207 selected participants. To minimize interrater variability and ensure annotation quality, an annotation guideline was developed under the supervision of neuroradiologists. The following guidelines were provided for areas prone to individual variations: (a) the area around the ventricles, sulci, and longitudinal fissure may show a non-WMH signal; therefore, an annotation is only made when a WMH is clearly visible on anterior to posterior slices, (b) the septum pellucidum of the lateral ventricles should not be annotated, (c) only WMHs in the cerebrum should be annotated, and (d) lacunar infarcts and EPVSs should be excluded. The imaging differentiation between lacunar infarcts, EPVSs, and WMHs is that a WMH is depicted as a clear hyperintense lesion on FLAIR, whereas an EPVS is depicted as an iso-to-hypointense lesion and a lacunar infarct is depicted as an iso-to-hyperintense lesion with a hypointense center [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In addition, WMHs are found mainly in the cerebral white matter, whereas EPVSs and lacunar infarcts are found in the basal ganglia, in addition to the cerebral white matter.\u003c/p\u003e \u003cp\u003eFor the model development, the annotated datasets were randomly divided into training (n\u0026thinsp;=\u0026thinsp;138) and test (n\u0026thinsp;=\u0026thinsp;69) datasets. To ensure an accurate evaluation, a second annotation review was conducted on 69 test participants. The review was performed by three neuroradiologists who differed from the five physicians employed in the initial annotation process. The reviewers were asked to modify the previously annotated WMH regions if needed. Consequently, some modifications were made to the test dataset.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eWMH Segmentation Challenge Dataset\u003c/h2\u003e \u003cp\u003eFor the evaluation using an external dataset, the WMHC Dataset was employed [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. A competition using this dataset was originally held at the 20th International Conference on Medical Image Computing and Computer-Assisted Intervention in 2017 in Quebec City, Quebec, Canada. After the competition, models could be submitted via their website until December 2022. Afterward, the previously undisclosed test dataset was made available. Several previous studies have used this dataset [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], and our study assessed performance using the same dataset. Although the dataset included both FLAIR and T1W images, only FLAIR images were used in this study.\u003c/p\u003e \u003cp\u003eThe WMHC dataset consisted of 60 training and 110 test participants. These participants were sourced from three institutes located in the Netherlands and Singapore. The data was acquired using five different types of scanners. The average slice thickness of the FLAIR sequences in this dataset was 2.23 mm. Notably the average slice thickness was approximately 41% that of the PR. The WMH volume averaged 16.9\u0026thinsp;\u0026plusmn;\u0026thinsp;21.6 mL, with a minimum of 0.78 mL and a maximum of 195.15 mL [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eWMH segmentation model\u003c/h2\u003e \u003cp\u003eThe algorithm for WMH detection in the software comprised a 2D convolutional neural network (CNN) model that processes individual MRI slices sequentially. Initially, pixel values of the input MRI images were normalized to fall within the range of 0.0\u0026ndash;1.0. Notably, image processing-based bias field corrections were not applied. We made this decision because the effect of the contrast augmentation was anticipated to accommodate such variations. Then, the normalized image was resized to 512 \u0026times; 512 pixels and input into the main CNN model. This CNN model was an ensemble of two U-Net models, each with distinct characteristics. The first U-Net model utilized ResNext50 as the encoder, and the ImageNet pretrained weights were used as the initial weights. The learning rate was set to 0.001, the batch size was set to 15, and the model was trained for 15 epochs using the Adam optimizer. These hyperparameters were obtained using a grid search. During training, the learning rate was decayed using a cosine-annealing learning-rate reducer to facilitate convergence. Matthews correlation coefficient loss was used as the loss function, as originally proposed for skin lesion segmentation [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. This loss function incorporates a penalty for misclassification of the true negative pixels and is expected to show robust performance. During the training process, data augmentations such as horizontal flip, vertical flips, shifting, rotation, contrast transformations, and grid distortion were applied to all the slices, regardless of WMH presence, to foster the development of a noise-robust model. The second U-Net model used EfficientNet-B5 as the encoder, which was trained in the same manner as the first U-Net. The output of the main CNN model was a threshold of 0.5, and the predicted binary WMH mask was obtained. The number of ensembles was set to two to obtain a balance between processing time and improved accuracy. Ensembles comprising three or more models increased the processing time by approximately 1.5 times or more, although the enhancement in performance was marginal.\u003c/p\u003e \u003cp\u003ePseudo-labeling was employed to effectively utilize the unlabeled data in the training of the actual model. First, a model for pseudo-labeling was trained using the PR for training only. The model was then used to predict the WMH regions in 885 unannotated participants. These predicted WMH masks can be regarded as pseudo-WMH annotations (pseudo-labeling dataset). Finally, a new model was trained using both the PR and pseudo-labeling dataset.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation metrics\u003c/h2\u003e \u003cp\u003eTo evaluate WMH segmentation performance, the voxel-wise DSC was used in this study. The DSC is a measure of the agreement between predicted and ground-truth regions and is defined as DSC\u0026thinsp;=\u0026thinsp;2 \u0026times; TP / (2 \u0026times; TP\u0026thinsp;+\u0026thinsp;FP\u0026thinsp;+\u0026thinsp;FN), where TP, FP, and FN are the numbers of true positive, FP, and FN voxels, respectively. Moreover, recall and precision were calculated using the following formulas: recall\u0026thinsp;=\u0026thinsp;TP / (TP\u0026thinsp;+\u0026thinsp;FN) and precision\u0026thinsp;=\u0026thinsp;TP / (TP\u0026thinsp;+\u0026thinsp;FP).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eProcessing time measurement\u003c/h2\u003e \u003cp\u003eTo evaluate the applicability of the model in clinical practice with limited resources, processing time was measured without using a graphics processing unit (GPU). The average of 10 measurements was computed for thick-slice MRI images. To facilitate comparison with a previous study [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e], the measured value was corrected for differences in the number of slices and central processing unit (CPU) clock frequencies as follows. In the prior study, time measurements were obtained using MRI with 192 slices on a computer with a 3.5-GHz CPU, whereas our study used 22 slices on a computer without a GPU, equipped with an Intel Core i5-10500T processor (Intel Corp., Santa Clara, CA, USA) running at 2.30 GHz with 16 GB of memory. Assuming that processing time decreases proportionally with the CPU clock frequency, the corrected processing time, intended for comparison with the results of the previous study, was calculated as follows: corrected processing time\u0026thinsp;=\u0026thinsp;measured time (min) \u0026times; (192/22) \u0026times; (2.3/3.5).\u003c/p\u003e \u003c/div\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e2D: two-dimensional, 3D: three-dimensional, AI: artificial intelligence, CNN: convolutional neural network, CPU: central processing unit, DSC: Dice similarity coefficient, DSWMH: deep and subcortical white matter hyperintensity, EPVS: enlarged perivascular space, FLAIR: fluid-attenuated inversion recovery, FN: false negative, FP: false positive, GPU: graphics processing unit, MRI: magnetic resonance imaging, PR: Private Dataset, PVH: periventricular hyperintensity, T1WI: T1-weighted image, T2WI: T2-weighted image, TP: true positive, WMH: white matter hyperintensity, WMHC: WMH Segmentation Challenge\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank all the radiologists and neurosurgeons for providing the diagnoses used in this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors have made substantial contributions to the intellectual content of the paper, contributed to data interpretation, approved the final manuscript, and agreed to its submission to this journal.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eM.K. contributed to the study design and concept; funding acquisition; research conduction; data collection, curation, management, and analysis; quality control; statistical analysis; and manuscript drafting.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eF.I. contributed to the study design and concept; funding acquisition; data collection and analysis; quality control; statistical analysis; and manuscript revision.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eS.N. and\u0026nbsp;S.K.\u0026nbsp;helped conceive and oversee the study and contributed to the revision of the manuscript.\u003c/p\u003e\n\u003cp\u003eD.I., H.K., T.H., and Y.M. contributed to data collection, curation, and analysis.\u003c/p\u003e\n\u003cp\u003eR.S., S.M., and T.K. helped with data collection and analysis and the revision of the manuscript.\u003c/p\u003e\n\u003cp\u003eY.S. helped conceive and oversee the study and assisted in data collection.\u003c/p\u003e\n\u003cp\u003eN.H. assisted in the planning and supervision of this study, as well as in the revision of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAnonymized data from the present study may be shared by the corresponding author upon request from a qualified researcher and upon permission from the institutional review board.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Japan Society for the Promotion of Science, Grant-in-Aid for Scientific Research, (C)23K08521.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eShinji Nakazawa, Ryo Sato, and Shiyuki Maeyama are employees of LPIXEL Inc. Yuki Shimahara owns stock in the company. Taiki Kaneko is currently an employee of Aile Home Clinic, Niigata, Japan. Masashi Kuwabara, Fusao Ikawa, Saori Koshino, Daizo Ishii, Hiroshi Kondo, Takeshi Hara, Yuyo Maeda, and Nobutaka Horie declare no potential conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWardlaw, J. M. \u003cem\u003eet al.\u003c/em\u003e Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration. Lancet Neurol. 12, 822\u0026ndash;838 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan Dijk, E. J. \u003cem\u003eet al.\u003c/em\u003e Progression of cerebral small vessel disease in relation to risk factors and cognitive consequences: Rotterdam Scan study. Stroke 39, 2712\u0026ndash;2719 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMosley, T. H. \u003cem\u003eet al.\u003c/em\u003e Cerebral MRI findings and cognitive functioning: the Atherosclerosis Risk in Communities study. Neurology 64, 2056\u0026ndash;2062 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDoddy, R. S., Massman, P. J., Mawad, M. \u0026amp; Nance, M. Cognitive consequences of subcortical magnetic resonance imaging changes in Alzheimer's disease: comparison to small vessel ischemic vascular dementia. Neuropsychiatry Neuropsychol. Behav. Neurol. 11, 191\u0026ndash;199 (1998).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eO'Brien, J. \u003cem\u003eet al.\u003c/em\u003e Severe deep white matter lesions and outcome in elderly patients with major depressive disorder: follow up study. BMJ 317, 982\u0026ndash;984 (1998).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTer Telgte, A. \u003cem\u003eet al.\u003c/em\u003e Cerebral small vessel disease: from a focal to a global perspective. Nat. Rev. Neurol. 14, 387\u0026ndash;398 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrins, N. D. \u0026amp; Scheltens, P. White matter hyperintensities, cognitive impairment and dementia: an update. Nat. Rev. Neurol. 11, 157\u0026ndash;165 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimoni, M. \u003cem\u003eet al.\u003c/em\u003e Age- and sex-specific rates of leukoaraiosis in TIA and stroke patients: population-based study. Neurology 79, 1215\u0026ndash;1222 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYamasaki, T. \u003cem\u003eet al.\u003c/em\u003e Prevalence and risk factors for brain white matter changes in young and middle-aged participants with Brain Dock (brain screening): a registry database study and literature review. Aging (Albany NY) 13, 9496\u0026ndash;9509 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreteler, M. M. \u003cem\u003eet al.\u003c/em\u003e Cognitive correlates of ventricular enlargement and cerebral white matter lesions on magnetic resonance imaging. The Rotterdam Study. Stroke 25, 1109\u0026ndash;1115 (1994).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNew Guidelines Development Committee for Brain Dock. [The Guideline for Brain Dock 2019]: Kyobunsha; 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMorita, A. Value of Brain Dock (Brain Screening) System in Japan. World Neurosurg. 127, 502 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u0026lt;vertical-align:sub;\u0026gt; \u0026lt;/vertical-align:sub;\u0026gt; Saito\u0026lt;vertical-align:sub;\u0026gt;,\u0026lt;/vertical-align:sub;\u0026gt; I. \u0026lt;background-color:#CCCCFF;vertical-align:sub;\u0026gt;[\u0026lt;/background-color:#CCCCFF;vertical-align:sub;\u0026gt;The Guideline for Brain Dock 2003\u0026lt;background-color:#CCCCFF;vertical-align:sub;\u0026gt;]\u0026lt;/background-color:#CCCCFF;vertical-align:sub;\u0026gt;\u0026lt;vertical-align:sub;\u0026gt;.\u0026lt;/vertical-align:sub;\u0026gt; Nihon Rinsho\u0026lt;vertical-align:sub;\u0026gt; \u0026lt;/vertical-align:sub;\u0026gt;64 Suppl 7\u0026lt;vertical-align:sub;\u0026gt;,\u0026lt;/vertical-align:sub;\u0026gt;\u0026lt;vertical-align:sub;\u0026gt; \u0026lt;/vertical-align:sub;\u0026gt;297\u0026ndash;302\u0026lt;vertical-align:sub;\u0026gt; \u0026lt;/vertical-align:sub;\u0026gt;\u0026lt;vertical-align:sub;\u0026gt;(\u0026lt;/vertical-align:sub;\u0026gt;\u0026lt;background-color:#66FF66;vertical-align:sub;\u0026gt;2006\u0026lt;/background-color:#66FF66;vertical-align:sub;\u0026gt;\u0026lt;vertical-align:sub;\u0026gt;)\u0026lt;/vertical-align:sub;\u0026gt;\u0026lt;vertical-align:sub;\u0026gt;.\u0026lt;/vertical-align:sub;\u0026gt;\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu, W. \u003cem\u003eet al.\u003c/em\u003e Automatic segmentation of white matter hyperintensities in routine clinical brain MRI by 2D VB-Net: a large-scale study. Front. Aging Neurosci. 14, 915009 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJoo, L. \u003cem\u003eet al\u003c/em\u003e. Diagnostic performance of deep learning-based automatic white matter hyperintensity segmentation for classification of the Fazekas scale and differentiation of subcortical vascular dementia. PLoS One 17, e0274562 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZijdenbos, A. P., Forghani, R. \u0026amp; Evans, A. C. Automatic \"pipeline\" analysis of 3-D MRI data for clinical trials: application to multiple sclerosis. IEEE Trans. Med. Imaging 21, 1280\u0026ndash;1291 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGrimaud, J. \u003cem\u003eet al.\u003c/em\u003e Quantification of MRI lesion load in multiple sclerosis: a comparison of three computer-assisted techniques. Magn. Reson. Imaging 14, 495\u0026ndash;505 (1996).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR\u0026oslash;vang M. S., \u003cem\u003eet al.\u003c/em\u003e Segmenting white matter hyperintensities on isotropic three-dimensional fluid attenuated inversion recovery magnetic resonance images: assessing deep learning tools on a Norwegian imaging database. PLoS One. 18, e0285683 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing, Y. \u003cem\u003eet al.\u003c/em\u003e Using deep convolutional neural networks for neonatal brain image segmentation. Front. Neurosci. 14, 207 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing, T. \u003cem\u003eet al.\u003c/em\u003e An improved algorithm of white matter hyperintensity detection in elderly adults. Neuroimage Clin. 25, 102151 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark, G., Hong, J., Duffy, B. A., Lee, J. M. \u0026amp; Kim, H. White matter hyperintensities segmentation using the ensemble U-Net with multi-scale highlighting foregrounds. Neuroimage 237, 118140 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, L., Kurgan, L., Wu, F. X. \u0026amp; Wang, J. Attention convolutional neural network for accurate segmentation and quantification of lesions in ischemic stroke disease. Med. Image Anal. 65, 101791 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLe M., \u003cem\u003eet al\u003c/em\u003e. FLAIR\u003csup\u003e2\u003c/sup\u003e improves LesionTOADS automatic segmentation of multiple sclerosis lesions in non-homogenized, multi-center, 2D clinical magnetic resonance images. Neuroimage Clin. 23, 101918 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeinen R, \u003cem\u003eet al\u003c/em\u003e. Performance of five automated white matter hyperintensity segmentation methods in a multicenter dataset. Sci. Rep. 9, 16742 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, \u003cem\u003eet al\u003c/em\u003e. A deep learning algorithm for white matter hyperintensity lesion detection and segmentation. Neuroradiology 64, 727\u0026ndash;734 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark, B. Y. \u003cem\u003eet al.\u003c/em\u003e DEWS (DEep White matter hyperintensity Segmentation framework): a fully automated pipeline for detecting small deep white matter hyperintensities in migraineurs. Neuroimage Clin. 18, 638\u0026ndash;647 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoeskops, P. \u003cem\u003eet al.\u003c/em\u003e Evaluation of a deep learning approach for the segmentation of brain tissues and white matter hyperintensities of presumed vascular origin in MRI. Neuroimage Clin. 17, 251\u0026ndash;262 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGibson, E., Gao, F., Black, S. E. \u0026amp; Lobaugh, N. J. Automatic segmentation of white matter hyperintensities in the elderly using FLAIR images at 3T. J. Magn. Reson. Imaging 31, 1311\u0026ndash;1322 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuijf, H. J. \u003cem\u003eet al.\u003c/em\u003e Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH Segmentation Challenge. IEEE Trans. Med. Imaging 38, 2556\u0026ndash;2568 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, H. \u003cem\u003eet al.\u003c/em\u003e Fully convolutional network ensembles for white matter hyperintensities segmentation in MR images. \u003cem\u003eNeuroimage\u003c/em\u003e 183, 650\u0026ndash;665 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRachmadi, M. F., Vald\u0026eacute;s-Hern\u0026aacute;ndez, M. D. C., Agan, M. L. F., Di Perri, C. \u0026amp; Komura, T. Segmentation of white matter hyperintensities using convolutional neural networks with global spatial information in routine clinical brain MRI with none or mild vascular pathology. Comput. Med. Imaging Graph 66, 28\u0026ndash;43 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTran, P. \u003cem\u003eet al.\u003c/em\u003e Automatic segmentation of white matter hyperintensities: validation and comparison with state-of-the-art methods on both Multiple Sclerosis and elderly subjects. Neuroimage Clin. 33, 102940 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFiford, C. M. \u003cem\u003eet al.\u003c/em\u003e Automated white matter hyperintensity segmentation using Bayesian model selection: assessment and correlations with cognitive change. Neuroinformatics 18, 429\u0026ndash;449 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, J., Zhang, Y. \u0026amp; Tang, X. Simultaneous tissue classification and lateral ventricle segmentation via a 2D U-net driven by a 3D fully convolutional neural network. \u003cem\u003eAnnu. Int. Conf. IEEE Eng. Med. Biol. Soc.\u003c/em\u003e 2019, 5928\u0026ndash;5931 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, D. \u003cem\u003eet al.\u003c/em\u003e Multi-atlas based detection and localization (MADL) for location-dependent quantification of white matter hyperintensities. Neuroimage Clin. 22, 101772 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLing, Y., Jouvent, E., Cousyn, L., Chabriat, H. \u0026amp; De Guio, F. Validation and optimization of BIANCA for the segmentation of extensive white matter hyperintensities. Neuroinformatics 16, 269\u0026ndash;281 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eManj\u0026oacute;n, J. V. \u003cem\u003eet al.\u003c/em\u003e MRI white matter lesion segmentation using an ensemble of neural networks and overcomplete patch-based voting. Comput. Med. Imaging Graph 69, 43\u0026ndash;51 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbhishek, K. \u0026amp; Hamarneh, G. Matthews correlation coefficient loss for deep convolutional networks: application to skin lesion segmentation. 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). \u003cem\u003eIEEE Xplore\u0026reg;\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://ieeexplore.ieee.org/document/9433782/authors#authors\u003c/span\u003e\u003cspan address=\"https://ieeexplore.ieee.org/document/9433782/authors#authors\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e, (2021).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-3833822/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3833822/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWe aimed to develop a new artificial intelligence software that can automatically extract and measure the volume of white matter hyperintensities (WMHs) in head magnetic resonance imaging (MRI) using only thick-slice fluid-attenuated inversion recovery (FLAIR) sequences from multiple centers. We enrolled 1,092 participants in Japan, comprising this thick-slice Private Dataset. Based on 207 randomly selected participants, neuroradiologists annotated WMHs using predefined guidelines. The annotated participants were divided into training (n\u0026thinsp;=\u0026thinsp;138) and test (n\u0026thinsp;=\u0026thinsp;69) datasets. The WMH segmentation model comprised a U-Net ensemble and was trained using the Private Dataset. Two other models were trained for validation using either both thin- and thick-slice MRI datasets or the thin-slice dataset alone. The voxel-wise Dice similarity coefficient (DSC) was used as the evaluation metric. The model trained using only thick-slice MRI showed a DSC of 0.820 for the test dataset, which is comparable to the accuracy of human readers. The model trained with the additional thin-slice dataset showed only a slightly improved DSC of 0.822. This automatic WMH segmentation model comprising a U-Net ensemble trained on a thick-slice FLAIR MRI dataset is a promising new method. Despite some limitations, this model may be applicable in clinical practice.\u003c/p\u003e","manuscriptTitle":"Artificial intelligence for volumetric measurement of cerebral white matter hyperintensities on thick-slice fluid-attenuated inversion recovery (FLAIR) magnetic resonance images from multiple centers","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-12 19:21:05","doi":"10.21203/rs.3.rs-3833822/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-03-04T05:44:30+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-02-28T19:00:51+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"d4c898b2-1f90-4bf4-870b-04adff0b6aba","date":"2024-02-19T18:17:12+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-02-12T12:28:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"79dc8d09-10e3-4042-9313-a6ea309a36ae","date":"2024-01-15T15:01:27+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-01-14T16:08:22+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-01-11T14:49:51+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-01-11T03:45:29+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-01-11T03:43:55+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-01-04T06:46:27+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"70d26d21-2fc1-4aea-b9e3-224d33f162b4","owner":[],"postedDate":"January 12th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":28088945,"name":"Biological sciences/Neuroscience/Cognitive ageing"},{"id":28088946,"name":"Physical sciences/Mathematics and computing/Software"},{"id":28088947,"name":"Health sciences/Anatomy/Nervous system"},{"id":28088948,"name":"Health sciences/Health care/Medical imaging/Magnetic resonance imaging"},{"id":28088949,"name":"Biological sciences/Neuroscience"},{"id":28088950,"name":"Health sciences/Medical research"},{"id":28088951,"name":"Health sciences/Neurology"},{"id":28088952,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2024-05-07T20:07:56+00:00","versionOfRecord":{"articleIdentity":"rs-3833822","link":"https://doi.org/10.1038/s41598-024-60789-x","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2024-05-02 19:57:53","publishedOnDateReadable":"May 2nd, 2024"},"versionCreatedAt":"2024-01-12 19:21:05","video":"","vorDoi":"10.1038/s41598-024-60789-x","vorDoiUrl":"https://doi.org/10.1038/s41598-024-60789-x","workflowStages":[]},"version":"v1","identity":"rs-3833822","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3833822","identity":"rs-3833822","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}