From our previous study using 3D T2-weighted MRI data acquired on MAGNETOM Skyra (Siemens AG, Munich, Germany) until September 2019 (Yamada et al., 2015, 2016a,b, 2017b, 2019), 14 patients (10 Hakim patients and 4 volunteers) were included in this study. Subsequently, from our recent study (Yamada et al., 2020, 2021c, 2023a,b,c), 115 patients (26 Hakim patients and 89 volunteers) who had undergone 3D T1-weighted and T2-weighted MRIs on a Discovery MR 750 W (GE Healthcare, Milwaukee, Wisconsin, United States) from October 2019 to January 2022, and 51 participants (6 Hakim patients and 45 volunteers) on a Signa Architect 3.0 T (GE Healthcare) from February 2022 to May 2022 were enrolled in this study. Healthy volunteers aged ≥20 years, were recruited from among medical staff, students, and their family members by open recruitment. The inclusion criteria for this study were as follows: individuals with no previous history of brain injury, brain tumor, or cerebrovascular disease on brain MRI examinations, and individuals who had never undergone brain CT or MRI and had no neurological symptoms, including compromised cognitive function. Three volunteers were incidentally detected with small unruptured intracranial aneurysms, but they were included in this study because small unruptured intracranial aneurysms were unlikely to affect brain and CSF volumes. One examination of 3D T1-weighted MRI in a healthy volunteer was excluded, because the MRI sequence and orientation differed completely from those of other images. Among 138 healthy volunteers, one was judged to have DESH, VD, and THC but not SFD, and was diagnosed with asymptomatic ventriculomegaly with features of iNPH on MRI (AVIM) (Iseki et al., 2009). All patients were diagnosed with or without Hakim disease, according to the third edition of the Japanese guidelines for management of iNPH (Nakajima et al., 2021). Among the 42 Hakim patients, 40 had triad symptoms of gait disturbance, cognitive impairment, and urinary incontinence, whereas two had very mild or no objective symptoms and did not undergo a CSF tap test or shunt surgery, and therefore would be classified as having AVIM. Overall, 18 patients (42%) underwent the CSF tap test, 21 patients (50%) underwent CSF shunt surgery, and their symptoms improved by ≥1 point on the modified Rankin Scale and/or the Japanese grading scale (Nakajima et al., 2021). Finally, 138 volunteers and 42 patients diagnosed with Hakim disease were included in this study (Table 1).

This study was approved by the ethics committees for human research at our institutes (IRB Number: 60-22-0083, R2019-227). Healthy volunteers provided written informed consent and underwent MRI examinations, after explaining the aim of this study and the potential for detection of diseases in the brain. Patients’ MRI data were obtained in an opt-out method, after their personal information was anonymized in a linkable manner.

The sequence parameters of T1-weighted 3D magnetization prepared rapid gradient echo (MPRAGE) were as follows: TR, 2471 ms; TE, 3.13 ms; inversion time, 1,000 ms; flip angle, 8°; matrix 256 × 256; voxel size, 0.9 × 0.9 × 0.9 mm; and acquisition time, approximately 4 min. The sequence parameters of 3D T2-weighted Cube were as follows: TR, 2000 ms; TE, 85.3 ms; matrix 288 × 288; voxel size, 0.8 × 0.8 × 0.8 mm; and acquisition time, approximately 4 min. The sequence parameters of 3D T2-weighted sampling perfection with application optimized contrast using the variable flip-angle evolution (SPACE) were as follows: TR, 2800 ms; TE, 286 ms; matrix 192 × 192; voxel size, 0.6 × 0.6 × 0.9 mm; and acquisition time, approximately 4 min.

As ground truth labels in our AI models, input image masks for volumetric semantic segmentation on the 3D T1-weighted MRI were created by combining manual segmentation with the 3D Viewer and fully automatic segmentation with the Brain Subregion Analysis applications (Figures 1A–C) on an independent 3D volume analyzer workstation (SYNAPSE 3D; FUJIFILM Corporation, Tokyo, Japan). In the Brain Subregion Analysis application, intracranial spaces were segmented fully automatically into 26 subregions including ventricles and SAS within 1 min (Yamada et al., 2023c). The input image masks from 3D T2-weighted MRI were also created using our original method, combining a simple threshold algorithm and manual segmentation (Figures 1D–F), as previously reported (Yamada et al., 2015, 2016a,b). Total SAS were further segmented into the Sylvian fissure and basal cistern, and the high-convexity SAS, which was defined as the location above the body of the lateral ventricles, with the lateral end 3 cm from the midline, the posterior end in the bilateral posterior parts of the callosomarginal sulci, and the anterior end on the coronal plane perpendicular to the AC–PC line passing through the front edge of the genu of the corpus callosum (Figure 2; Supplementary Videos S1–S4) (Yamada et al., 2023a). All input image masks as the ground truth labels were transferred to the SYNAPSE Creative Space for cloud-based AI development service (FUJIFILM Corporation). All masks were processed and formatted into a form that could utilize the training or inference process. Regarding the output of the inference process, feature maps were obtained. Overall, 159 T1-weighted images were assigned to 110 images for training, 30 for internal and 19 for external validation (test), and 180 T2-weighted images were assigned to 130 images for training, 30 for internal validation and 20 for external validation. Inference was performed in the images for internal validation and external validation.

We combined two deep learning models to employ a two-step method of automatic detection of DESH with segmented volumes and indices. In the first step, the volumetric semantic segmentation task employed a 3D U-Net with four layers, consisting of 3D convolution with a batch normalization layer, ReLU activation layer, max pooling layer, and 3D up-convolution layer (Figure 3A). Signal values were normalized by percentile (minimum 0.05, maximum 0.95) as a preprocessing step. To compensate for voxel detail, feature maps are concatenated from each encoding layer of feature extraction by downsampling to the corresponding decoding layer of feature assignment by upsampling. In the second step, the image classification task employed a multimodal convolutional neural network (CNN) (Figure 3B). As ground truth labels for the image classification task, the presence or absence of DESH, VD, THC, and SFD was determined by a neurosurgeon and a radiologist, both experts in imaging diagnosis of Hakim disease, through consensus reading. Input data included the presence of DESH, VD, THC, and SFD, in addition to age at MRI, gender, and the same image masks used in the first step volumetric semantic segmentation task (Figure 3B). For the output of the image classification task, the intracranial CSF space mask was used to determine the presence or absence of DESH, and the masks for the total ventricle, high-convexity SAS, and Sylvian fissure and basal cistern were used to determine the presence of VD, THC, and SFD, respectively. In the embedding layer, all input variables were transformed into feature maps. At the end of the last convolutional layer, the final feature maps were fed to a softmax activation function to generate a probability score for each class. Image intensities of input images were normalized to [0, 1] by their maximum and minimum values. Augmentations including rotation, scaling, and translation of the input image masks were made to improve generalizability and accuracy in the semantic segmentation and image classification tasks. The generalizability of these augmentations would help reduce effects from differences between manufacturers, imaging protocols or individuals, and increase the robustness of our AI model.

The “DESH index” was defined as the combined volume of total ventricles and Sylvian fissure and basal cistern divided by the high-convexity SAS volume. As supplemental indices for DESH, the “Venthi index” was defined as the total ventricular volume divided by the high-convexity SAS volume, and the “Sylhi index” was defined as the volume of the Sylvian fissure and basal cistern divided by the high-convexity SAS volume. These three indices were calculated by the manually and automatically segmented volumes.

Mean age and segmented volumes were compared using the Mann–Whitney-Wilcoxon test. The chi-square test was used to compare the proportions between Hakim patients and healthy volunteers. To quantify the performance, e.g., the accuracy of the volumetric semantic segmentation, the Dice coefficient score for the loss function was calculated as 2 * |X ∩ Y| + epsilon(1e-4)/(|X| + |Y| + epsilon(1e-4)) in the validation study. X and Y were the prediction and correct, binary [0, 1] output per voxel. The relationships between the manually and automatically segmented volumes were also examined using Pearson’s correlation coefficient (r) and 95% confidential intervals (CIs). For the image classification task, the accuracy and softmax probability score for the detection of DESH, VD, THC, and SFD were analyzed. The area under the receiver-operating characteristic curves (AUCs) and optimal thresholds with 95% CIs for detecting DESH were calculated to maximize the sum of the sensitivities and specificities. All missing variables were considered deficit data, and no other variables were adjusted. A probability value (P) of <0.001 was considered to be statistically significant. R software (version 4.2.1, R Foundation for Statistical Computing, Vienna, Austria, http://www.R-project.org) was used for all statistical analyses.

We prepared 180 datasets of 3D T2-weighted MRIs and 159 datasets of 3D T1-weighted MRIs. All 3D T1-weighted MRIs were MPRAGE sequence, and 166 3D T2-weighted MRIs were Cube sequence and 14 were SPACE sequence. For both deep learning models, 110 T1- and 130 T2-weighted MRIs were used for training, 30 T1- and 30 T2-weighted MRIs for internal validation, and the remaining 19 T1- and 20 T2-weighted MRIs for external validation. The allocation of the number of DESH or non-DESH is shown in Table 2.

Training and internal validation of the 3D U-Net model for semantic segmentation were repeated over 1,000 times (Figures 4–7; Supplementary Figures S1, S2). Overall, the intracranial CSF space, total ventricles, total SAS, Sylvian fissure and basal cistern, and the high-convexity SAS were segmented fully automatically from 3D T1-weighted (Figure 8) and T2-weighted MRIs (Figure 9). There was no significant difference between manually and automatically segmented volumes of the total ventricles, total SAS, high-convexity SAS, and Sylvian fissure and basal cistern (Table 3). Among the segmented regions, the mean Dice scores for the total ventricles were highest (0.85 from T1 and 0.83 from T2), those for the Sylvian fissure and basal cistern were second highest (0.70 and 0.69), and those for the high-convexity SAS were lowest (0.68 and 0.60). The mean Dice coefficient scores for all of the regions segmented from the T1-weighted image were superior to those from the T2-weighted image. The mean differences between the manually and automatically segmented volumes of the high-convexity SAS were smaller (T1 and T2; 3.6 mL and 4.2 mL) than those of the Sylvian fissure and basal cistern (5.3 mL and 8.3 mL).

The inference results of the presence or absence of DESH, VD, THC, and SFD with softmax probability scores are summarized in Table 4. All mean softmax probability scores were exceeded by 0.99, except for THC detection from the T1-weighted image (0.95) and SFD detection from the T2-weighted image (0.98). Among 99 images (49 T1 and 50 T2), only one T1-weighted image of a volunteer was judged by AI to have DESH, but the expert judged the subject to have no DESH (Figure 10). In addition, the discrepancy between the AI and expert determinations from T1-weighted MRIs was one for VD, one for THC, and three for SFD. However, AI determinations from T2-weighted MRIs were almost perfectly consistent with expert determinations, with only one discrepancy in VD determination. The accuracies for the determinations of DESH, VD, THC, and SFD by AI were 1.0, 1.0, 1.0, and 0.97 from T1-weighted MRIs, and 1.0, 1.0, 1.0, and 0.93 from T2-weighted MRIs, respectively.

All DESH, Venthi and Sylhi indices, calculated by the manually and automatically segmented volumes on T1-weighted and T2-weighted MRIs, had sufficiently high AUCs (>0.996), specificities (>0.944), and sensitivities (>0.923) (Table 5). However, optimal thresholds calculated to maximize the sum of sensitivities and specificities for detecting DESH differed between manually and automatically segmented volumes and between T1-weighted and T2-weighted MRIs.